Nlsam

- Support for numpy 2.0 and python 3.9 and up
- Fixes for cython 3 and newer scipy

- Some speed improvements internally
- Some more functions in parallel
- A new progress bar with tqdm
- New non-frozen builds for the standalone versions
- Mac M1/M2 arm64 binary wheels now available

- **Breaking changes in the command line parser**
- The previously required options __N__ and __angular_block_size__ are now optional.
- A mask is now required to be passed with __-m__ or __--mask__ to only sample data. It was previously possible to be unlucky and only sample background noise in the reconstruction process, taking forever to practically do nothing in practice, passing a mask with only the data to sample and reconstruct should prevent this issue.

- A new call would now looks like

~~~bash
nlsam_denoising input output bvals bvecs -m mask.nii.gz
~~~

- New command line arguments, now subclassed into categories.
+ __--load_mhat__ file, to load a volume for initializing the bias correction, the default is to use the data itself.
+ __--save_difference__ file, to save a volume showing the removed signal parts as abs(original_data - denoised_data)
+ __--save_eta__ file, to save the result of the estimated underlying signal value for debugging purposes.
+ Deprecation of options __--implausible_signal_fix__ and __--sh_order__, use __--load_mhat__ instead for initialization.

- Support for non-integer values of N.
- Support for supplying a volume to be loaded as values of N.
- New module nlsam.bias_correction, which contains an easier to use interface to the C functions in nlsam.stabilizer
- New online documentation available at http://nlsam.readthedocs.io/ for the current (and future) versions.
- The dictionary learning part of the algorithm now respects **--cores** instead of ignoring it and always using all available processors.
- joblib is now used for parallel processing.
- The frozen executable is now using dask and performs a bit slower than the normal version until joblib.loky is fixed to work with pyinstaller.
- Binary wheels are now available for all platforms instead.
- A new option to estimate automatically the noise distribution (sigma and N) is now available by passing **auto** to both N and **--noise_est**.
- This option is also the new default now.
- A new option to process each shell separately is now available with **--split_shell**.
- Probably other stuff I forgot.

- Fixed a numerical issue in the Marcum Q function when computing probabilities used in the stabilizer framework.
- Scipy >= 0.19.1 is now required.
- nlsam.stabilizer.stabilization now accepts the keyword clip_eta (default True), which can be used to allow returning negatives values for eta.
- The option __--no_clip_eta__ from nlsam_denoising can be used to activate this feature.
- The previous versions forced negative values to zero and is still the default behavior.

- PIESNO will now warn if less than 1% of noisy voxels were identified, which might indicate that something has gone wrong during the noise estimation.
- On python >= 3.4, __--mp_method__ [a_valid_start_method](https://docs.python.org/3/library/multiprocessing.html#contexts-and-start-methods) can now be used to control behavior in the multiprocessing loop.
- A new option __--split_b0s__ can be specified to split the b0s equally amongst the training data.
- A new (kind of experimental) option __--use_f32__ can be specified to use the float32 mode of spams and reduce ram usage.
- A new option __--use_threading__ can be specified to disable python multiprocessing and solely rely on threading capabilities of the linear algebra libs during denoising.
- Fixed crash in option __--noise_est__ local_std when __--cores__ 1 was also supplied.
- setup.py and requirements.txt will now fetch spams v2.6, with patches for numpy 1.12 support.
- The GSL library and associated headers are now bundled for all platforms.
- Some deprecation fixes and other performance improvements.

- Fixed a bug in local piesno (option __--noise_map__) where the noise would be underestimated.
- Fixed a bug introduced in v0.5 where datasets with multiple b0s would be incorrectly reshaped and dropped some volumes. Thanks to Samuel Deslauriers-Gauthier for reporting.