Refactor parallelism to support PyTorch and only use shared memory

[mdekstrand]

This reworks LensKit's parallel processing support to use PyTorch's shared memory capabilities instead of custom shm buffers or binpickle, and moves it to a new lenskit.parallel package. Further work on #374. Closes #382.

[codecov]

Codecov Report

Attention: Patch coverage is 93.20000% with 17 lines in your changes are missing coverage. Please review.

Project coverage is 91.68%. Comparing base (d940076) to head (c4c68ab).

Files	Patch %	Lines
lenskit/parallel/config.py	91.07%	5 Missing ⚠️
lenskit/parallel/pool.py	93.61%	3 Missing ⚠️
lenskit/parallel/serialize.py	93.61%	3 Missing ⚠️
lenskit/parallel/worker.py	92.10%	3 Missing ⚠️
lenskit/parallel/invoker.py	92.59%	2 Missing ⚠️
lenskit/batch/_recommend.py	85.71%	1 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main     #393      +/-   ##
==========================================
- Coverage   92.25%   91.68%   -0.57%     
==========================================
  Files          41       43       +2     
  Lines        3162     3031     -131     
==========================================
- Hits         2917     2779     -138     
- Misses        245      252       +7     

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