start fixing metric tests

[ivirshup]

• Closes Metrics started returning inconsistent results #2688

• Release notes not necessary because: just modifying tests

Fixes tests for metrics

Some notes on an in progress PR:

• Previously xfail tests didn't actually fail because nothing was asserted
• This behavior changes with version of numba.
  • numba .56<= seems more reproducible, but differences are greater when they occur (e.g. calculating on sparse vs dense)
• Ideally want per metric, per calculation tolerances
• Both threading options can differer
• I think single threaded + fastmath=False is reproducible, but need to confirm – still no

[codecov]

Codecov Report

All modified and coverable lines are covered by tests ✅

Comparison is base (bf5f27a) 72.72% compared to head (28c2def) 72.72%.
Report is 1 commits behind head on master.

Additional details and impacted files

@@           Coverage Diff           @@
##           master    #2705   +/-   ##
=======================================
  Coverage   72.72%   72.72%           
=======================================
  Files         111      111           
  Lines       12383    12383           
=======================================
  Hits         9005     9005           
  Misses       3378     3378           

Files	Coverage Δ
scanpy/plotting/_utils.py	56.33% <100.00%> (ø)