Bam group by speedup

[Lioscro]

Using hits.utilities.group_by happens to be very slow for some reason, so I implemented our own group_bam_by_key function, which achieves approx. 100x speedup.

Branches from remove-intbc-logging #163

[codecov]

Codecov Report

Patch coverage: 93.87% and no project coverage change.

Comparison is base (013f615) 79.53% compared to head (fe28c8b) 79.54%.

Additional details and impacted files

@@           Coverage Diff           @@
##           master     #164   +/-   ##
=======================================
  Coverage   79.53%   79.54%           
=======================================
  Files          89       89           
  Lines        7931     7948   +17     
=======================================
+ Hits         6308     6322   +14     
- Misses       1623     1626    +3     

Files Changed	Coverage Δ
cassiopeia/preprocess/cassiopeia_preprocess.py	21.81% <0.00%> (-1.26%)	⬇️
cassiopeia/preprocess/pipeline.py	91.15% <ø> (ø)
cassiopeia/preprocess/UMI_utils.py	94.71% <100.00%> (+0.34%)	⬆️

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[colganwi]

For me this PR does not speedup cas.pp.collapse at all. @Lioscro can you provide a test case or more details about the compute environment? Maybe this is faster with memory constraints?

[Lioscro]

It's been a hot minute since I've worked on this, so I'm having a hard time providing any additional context, but I believe the speedup was achieved for very large BAMs.

But given my limited recollection, I'd trust your experience more than whatever I can remember, so I'm fine if you decide this PR is no longer needed.

[colganwi]

I tested with a 15GB bam file and see a 2X speed up. MD5 checksums of the outputs are the same. Nice work @Lioscro!

[colganwi]

Looks good to me!