Preserve prompt padding masks during conditioning

[zjn20030811]

Summary

The SimWAM prompt-conditioning path currently replaces the tokenizer validity
mask with an all-ones mask after zeroing padded embeddings. As a result,
padded rows can still participate in cross-attention (for example after a
projection bias or in attention implementations that consume the mask).

This change preserves the tokenizer mask and defensively zeros invalid rows
with an out-of-place operation that also works for non-contiguous encoder
outputs.

Validation

• pytest -q tests/test_prompt_mask.py (2 passed)
• python -m compileall -q src/simwam tests
• git diff --check

The regression tests cover both the returned prompt mask/embeddings and the
invariance of cross-attention to changes in padded context rows.