fix: defer bitsandbytes import until use

[motodriver]

Summary

• remove the unconditional module-level bitsandbytes import
• import bitsandbytes only when a quantized weight actually requires 4-bit dequantization
• add a subprocess regression test that rejects any bitsandbytes import while loading heretic.model

This prevents an unrelated bitsandbytes native-library traceback from being emitted at startup for non-BNB use while preserving the existing failure behavior when 4-bit dequantization is actually required.

Fixes #429.

Verification

• uv run --frozen python -m unittest discover -s tests -p "test_*.py" — 5 passed
• uv run --frozen ruff format --check .
• uv run --frozen ruff check --output-format=github --extend-select I .
• uv run --frozen ty check --output-format=github --error-on-warning .
• git diff --check

AI assistance from TRAE was used for repository analysis, test drafting, and review. I reviewed the complete diff and ran the verification above.

[p-e-w]

Get rid of this test, it doesn't tell us anything. It simply asserts that the import line that has been moved has indeed been moved.