Fix Linear .nam tap serialization order

[sdatkinson]

Summary

• reverse the convolution tap axis when exporting Linear models so .nam weights use chronological impulse-response order
• reverse serialized taps when importing them into PyTorch while leaving bias placement unchanged
• cover asymmetric impulses, imports, bias-preserving round trips, and Linear children inside Sequential models

Tests

• python -m pytest -q tests/test_nam (353 passed, 21 skipped)
• python -m pytest tests/test_nam/test_models/test_linear.py tests/test_nam/test_models/test_sequential.py tests/test_nam/test_models/test_from_nam.py -q
• python -m flake8 . --count --select=E9,F63,F7,F82 --show-source --statistics
• Full-suite CLI integration tests remain environment-blocked because the globally installed nam-full entry point resolves to a different checkout and fails before training with ModuleNotFoundError: No module named 'nam.train.full'.

Resolves #696