Fix SoloSeq embedding truncation behavior

[samanyuckulkarni123]

Summary

Fix SoloSeq embedding truncation behavior in scripts/precompute_embeddings.py.

The previous implementation truncated with toks[:1022], which slices the batch axis rather than the token-length axis. Fixes that bug and makes long sequence handling explicit.

Details

• truncate ESM inputs on the token-length axis instead of the batch axis
• preserve saved embedding length based on the effective truncated sequence length
• replace the old truncate flag behavior with a real --truncate / --no-truncate interface
• emit a warning when sequences longer than 1022 residues are truncated
• raise a clear error when truncation is disabled for overlength sequences
• add focused regression tests
• update SoloSeq docs to match the implemented behavior

Test plan

• Verified the updated files parse successfully:
  • python3 -m py_compile scripts/precompute_embeddings.py tests/test_precompute_embeddings.py
• Ran the new targeted regression test module in a local virtual environment with torch installed:
  • .venv/bin/python -m unittest tests.test_precompute_embeddings
• Confirmed the test module covers:
  • default CLI behavior (--truncate defaults to enabled)
  • explicit --no-truncate behavior
  • correct truncation of overlength sequences to the ESM-1b 1022 residue limit
  • truncation on the token-length axis rather than dropping batch rows
  • clear failure when truncation is disabled for overlength input