Use upstream sklearn PCA if possible

[flying-sheep]

this PR fixes #2744 and

• moves _pca_with_sparse behind a check for scipy <1.4, which has @ivirshup’s port of that code ENH Allow fitting PCA on sparse X with arpack solvers scikit-learn/scikit-learn#18689

• simplifies our logic around which parameters lead to which dispatch

  this makes it useful to get this in before Implement sparse covariance_eigh PCA using Dask #3263

• throws a warning when people use the lobpcg solver, since the closure of PCA on sparse, noncentered data scikit-learn/scikit-learn#12794 (comment) makes it unlikely that we can count on that getting in any time soon

  I filed Add svd_solver="lobpcg" option to PCA scikit-learn/scikit-learn#30075. Depending on how that PR is received, we can update the warning here: either they like it, then we can remove the warning (we remove our legacy code once we depend on a scipy version that has lobpcg upstream), or they don’t, then we leave the warning for now and remove lobpcg support in the future.

[codecov]

Codecov Report

Attention: Patch coverage is 95.16129% with 6 lines in your changes missing coverage. Please review.

Project coverage is 77.03%. Comparing base (bbcd4b1) to head (7516acc).
Report is 59 commits behind head on main.

Files with missing lines	Patch %	Lines
src/scanpy/preprocessing/_pca/__init__.py	93.18%	6 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #3267      +/-   ##
==========================================
+ Coverage   76.95%   77.03%   +0.08%     
==========================================
  Files         109      110       +1     
  Lines       12465    12492      +27     
==========================================
+ Hits         9592     9623      +31     
+ Misses       2873     2869       -4     

Files with missing lines	Coverage Δ
src/scanpy/preprocessing/_pca/_compat.py	100.00% <100.00%> (ø)
src/scanpy/preprocessing/_simple.py	90.64% <ø> (+2.00%)	⬆️
src/scanpy/preprocessing/_pca/__init__.py	91.20% <93.18%> (ø)

[ilan-gold]

Some comments and questions, some changes needed. Overall, much cleaner :)