fix(openvla): match action token decode to OpenVLA bin centers

[rylinjames]

Summary

• Decode OpenVLA action tokens against the effective text vocab size used by the HuggingFace model, defaulting to 32000 for openvla-7b instead of the padded LM size.
• Map action bins through centers between the configured action edges, matching OpenVLA's bin center postprocess.
• Update shim docs and tests to cover center-bin behavior while preserving padded-logit output shapes.

Validation

• PYTHONPATH=$PWD/src /Users/romirjain/Desktop/building\ projects/fastcrest/tether/.venv/bin/python -m pytest tests/test_openvla_postprocess.py tests/test_openvla_shim.py -p no:cacheprovider
• PYTHONPATH=$PWD/src /Users/romirjain/Desktop/building\ projects/fastcrest/tether/.venv/bin/ruff check src/tether/postprocess/openvla.py src/tether/exporters/openvla.py tests/test_openvla_postprocess.py
• PYTHONPATH=$PWD/src /Users/romirjain/Desktop/building\ projects/fastcrest/tether/.venv/bin/python -m py_compile src/tether/postprocess/openvla.py src/tether/exporters/openvla.py

🤖 Generated with Claude Code