Pp-sketchlib

Small improvement to poor regressions, to fix a CI failure in PopPUNK

What's Changed
* Fix buildkite by johnlees in https://github.com/bacpop/pp-sketchlib/pull/87
* More robust regression with zero distances by johnlees in https://github.com/bacpop/pp-sketchlib/pull/88


**Full Changelog**: https://github.com/bacpop/pp-sketchlib/compare/v2.1.0...v2.1.1

- Remove armadillo from the code entirely in favour of Eigen
- Actually link BLAS (used in the regression), LAPACK not actually needed
- Use a new regression solver, see 'Normal equations' under https://eigen.tuxfamily.org/dox/group__LeastSquares.html
- Solve issues with `bool` in highfive code, now uint8_t explicitly used
- Use KMeansRex, as the kmeans in armadillo is now removed

nvcc now sees the eigen code and throws up a few warnings, but this doesn't seem to be causing as many problems as it used to

What's Changed
* BLAS linking; CI fix; better message for deprecated function by johnlees in https://github.com/bacpop/pp-sketchlib/pull/84

**Full Changelog**: https://github.com/bacpop/pp-sketchlib/compare/v2.0.1...v2.1.0

Minor update which should not change/break the API, adds a few small new features. 82 simplifies and parallelises sparsification by a threshold (and removes kNN).

What's Changed
* Change web sketch output from uint64 to hex by muppi1993 in https://github.com/bacpop/pp-sketchlib/pull/74
* Update README.md by danrlu in https://github.com/bacpop/pp-sketchlib/pull/75
* Add reads attribute to databases by johnlees in https://github.com/bacpop/pp-sketchlib/pull/77
* Lineage model fitting - sketchlib changes by nickjcroucher in https://github.com/bacpop/pp-sketchlib/pull/82

New Contributors
* muppi1993 made their first contribution in https://github.com/bacpop/pp-sketchlib/pull/74
* danrlu made their first contribution in https://github.com/bacpop/pp-sketchlib/pull/75

**Full Changelog**: https://github.com/bacpop/pp-sketchlib/compare/v2.0.0...v2.0.1

This is a new major version due to overhaul of the CLI.

New features:
- CLI is now written with docopt, and should be more straightforward to use (72).
- New self sparse query, which works with CUDA too (67). Used in [mandrake](https://github.com/bacpop/mandrake).

Bug fixes:
- Sparsification functions fixed and added to tests.

Other:
- Enable pip install and pushing to pypi (70)
- Adding docker build and test on buildkite (70)

New features:
- Sketching kernel overhaul: should be faster, and supports read datasets larger than the device memory size (63 64)
- Increase the size of the countmin table on GPUs to 2^30 columns, takes up about 30% of a 3090's memory (64)
- Better progress meter for read sketching (64)
- Save whether the DB was calculated using the reverse complement strand (65)

Bug fixes:
- Fix indices for reference vs query distance calculations, hit when reaching ~100k queries (64)
- If sketching craps out early, save the successful sketches up until that point to the HDF5 DB without corrupting it, so it can be joined later. (64)

Other:
- Refactor the CUDA code a bit, putting classes into their own files. (64)

New features:
- Better parallel efficiency for distance calculation on a CPU (62)

Bug fixes:
- Progress meter caused crash with small numbers of distances (62)
- Smoother progress meter for distances (62)
- Remove device reset which could cause problems with both GPU dists and a GPU graph in poppunk (62)