Refactor HMM inference and training backends

[AKuederle]

Summary

• replace the pomegranate-bound HMM object graph with a backend-neutral HmmModel containing plain NumPy parameters
• preserve the pretrained Roth model's exact legacy hidden-state and stride-list inference through ScipyHmmInference
• add modern pomegranate 1.x training and inference backends for Python 3.10+, plus explicit legacy pomegranate inference on Python 3.9
• make Roth training, inference, serialization, and tpcp optimization use the new backend seams and remove the old HMM implementation
• load old Roth/pomegranate JSON exports into the compiled model so packaged and custom pretrained models work with every inference backend

Verification

• full local suite: 2432 passed, 37 skipped
• Python 3.9 isolated HMM and public-boundary tests
• clean Sphinx builds on Python 3.9 and Python 3.11
• GitHub Actions: lint and all Python 3.9/3.10/3.11 Ubuntu/Windows jobs passed
• Read the Docs build 33625193 succeeded
• final roborev review passed with no findings; no review jobs remain open

[codecov-commenter]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 88.17%. Comparing base (05ad971) to head (ad2f788).

Additional details and impacted files

@@           Coverage Diff           @@
##           master      #85   +/-   ##
=======================================
  Coverage   88.16%   88.17%           
=======================================
  Files          62       62           
  Lines        3575     3561   -14     
=======================================
- Hits         3152     3140   -12     
+ Misses        423      421    -2     

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

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.