Banditpam

**Organization and Functionality:**
- Added a new GHA to build the package and run tests on Windows.
- Added a configuration file v2 for our Read the Docs documentation.
- Added a convenience script `retrieve_windows_python_files.sh` to retrieve `clang_rt.asan_dynamic-x86_64.dll` automatically for the Windows Python build.

**Tests:**
- Updated `tests/test_smaller.py` to use different data types on different platforms when loading `data/scrna_reformat.csv.gz` to adjust memory usage.

**Documentation:**
- Updated the Windows installation instructions to use the convenience script `retrieve_windows_python_files.sh`.

**Full Changelog**: https://github.com/motiwari/BanditPAM/compare/v4.0.3...v4.0.4

**Organization and Functionality:**
- Ensure GHAs don't run on a tag push.
- Automatically upload wheels to TestPyPI on a PR update and PyPI on release in GHAs (fixes 256)
- Add `tests/test_initialization` to the GHAs that build and test the package.
- Loss issues are resolved in all versions of BanditPAM
- The clustering results are now identical to `scikit-learn`'s implementation. This was achieved by increasing the batch size for accurate estimation of the standard deviation of the arm parameters and by fixing the bug that drops arms whose lower confidence bounds equal the lowest upper confidence bounds (fixes 252).
- Add complexity to `scripts/comparison_utils.py`, such as printing cache writes.
- Update cache calculations in `src/algorithms/kmedoids_algorithm.cpp` and add an assertion for better error handling.
- Turn `VERSION_INFO` into a string in `src/python_bindings/kmedoids_pywrapper.cpp`.
- Update `build_confidence` in `src/python/kmedoids_pywrapper.cpp` to avoid stochasticity issue in `tests/test_larger.py`.

**Tests:**
- Update assertion in `tests/test_initialization` to match with new `build_confidence` in `src/python/kmedoids_pywrapper.cpp`.
- Use NumPy and Armadillo seeds for reproducibility.
- Update some tests in `tests/test_larger.py` to count the proportion of passing cases instead of immediately failing.

**Style:**
- Ran `black`, `docformatter`, `flake8`, and `clang-format`.

**Documentation:**
- Slightly updated the documentation on the CMake build on Windows to clarify how to use `retrieve_windows_cmake_files.sh`.

**Full Changelog**: https://github.com/motiwari/BanditPAM/compare/v4.0.2...v4.0.3

**Organization and Functionality:**
- We include the files, edits, and instructions to succeed in the CMake and Python builds on Windows, Mac, and Linux.
- Note that `carma` has now bumped to [version v0.6.7](https://github.com/RUrlus/carma/commit/8405bafc40b1bf06795e7004c7898a08963b688c).
- Does not require LLVM's clang when running `pip install .` (Fixes 242)
- Swaps are only performed in `banditpam.cpp`, `banditpam_orig.cpp`, `fastpam1.cpp`, and `pam.cpp` if k=1 (Fixes 227)
- Uses `bool` data type when necessary (Fixes 232)
- Implements the `buildLoss` functionality (Fixes 234)
- Modifies the FastPAM1 implementation such that the runtime is significantly reduced while resulting in the same medoids (Fixes 228)
- Ensures Linux GHA error does not appear anymore (Fixes 250)
- Update GHAs to use `python -m pip` throughout instead of just `pip`
- Mac GHAs now run on `macos-latest` as `macos-10.15` is now unsupported.
- Fixes failing GHAs that activated on PR.

**Tests:**
- Updated incorrect assertion in `tests/test_initialization.py`

**Style:**
- Ran `black`, `flake8`, a combination of `clang-format` and `cpplint`, and our own style definitions over codebase (such as using 2 spaces instead of 4 in C++ code).

**Documentation:**
- Documented functional CMake and Python builds on Windows.

**Full Changelog**: https://github.com/motiwari/BanditPAM/compare/v4.0.1...v4.0.2

**Organization and Functionality:**
- Makes `cibuildwheel` more verbose
- For CMake builds, checks out the current branch (not `main`) in GHA
- Changes the CMake build to `Release` mode to avoid running sanitizers in GHA
- Adds status badges (Fixes 224)
- Fixes the GHA builds by including the proper directory for `armadillo` (the locally installed version)

**Tests:**
No changes.

**Style:**
No changes.

**Documentation:**
- Updates `README.md` and other docs

**Organization and Functionality:**
- We now rely on C++17 for `std::optional`
- We implement a new version of BanditPAM that considers only $n$ non-medoids as arms in the SWAP step, each of which has $k$ values. This leads to a significant speedup. The old BanditPAM is still available as `BanditPAM_orig`
- Update the GHA to use newer machine images and update the dependencies on Ubuntu 22.04
- Split the cmake build into a separate GHA
- Added Python 3.11 to GHA where possible, removed Python 3.6
- Restores the ability to call OpenMP functions `omp_get_max_threads` and `omp_set_num_threads`, which should resolve the remaining issues on M1 Macs
- We allow users to pass in a distance matrix (Fixes 164 )
- Now uses `buildConfidence` and `swapConfidence` in log (ln) space
- Allows users to retrieve the sample complexities and caching statistics
- Adds `scripts/cache_measurements.py`, `scripts/compare_banditpam_versions.py`, `scripts/comparison_utils.py`, `scripts/comparison_with_fasterpam.py`, `scripts/dist_mat_test.py`, `scripts/timing_dist_mat.py`, `scripts/sample_complexity_with_k.py`, and `scripts/scaling_with_k.py`
- Properly builds Mac wheels using Apple Clang and enabling OpenMP. Tested across both Intel and M1 Macs for `gcc`- and `clang`- compiled Python.
- Adding `parallelize` flag that only uses OpenMP parallelization if set to `true`, i.e., replacing all `pragma omp parallel for` with `pragma omp parallel for if (this->parallelize)`
- Fixing an issue with the python bindings where variable orderings didn't match so they were getting intialized with each other's values
- Changes `max_iter` to `100`
- Added `tests/test_initialization.py`, but this is not run by GHA
- Implements caching of distance computations. Also allows for retrieving cache statistics (hits, misses, writes)
- Allows setting cache parameters (`useCache`, `usePerm`, etc.) from Python
- Allows setting `parallelize` from Python

**Tests:**
- Added `tests/test_initialization.py`, but this is not run by GHA

**Style:**
- Ran `black` over codebase.

**Documentation:**
- Documented new capabilities.

**Full Changelog**: https://github.com/motiwari/BanditPAM/compare/v3.0.4...v4.0.0

**Organization and Functionality:**
- Fixes the computation of cosine distance (Fixes 182)
- Removes the ability to call OpenMP functions `omp_get_max_threads` and `omp_set_num_threads`, which should resolve the remaining issues on M1 Macs (Fixes 167)

**Tests:**
No changes.

**Style:**
No changes.

**Documentation:**
No changes.

**Full Changelog**: https://github.com/ThrunGroup/BanditPAM/compare/v3.0.3...v3.0.4