Fix NonCausalAttnPress key-padding mask value 🤖🤖🤖

[AmirF194]

PR description

non_causal_chunked_attn masks padded key columns in the ragged last chunk with -1e-9 before the softmax (kvpress/presses/non_causal_attention_press.py:91). -1e-9 is close to zero, so it barely suppresses the padded columns: a real query in the last chunk leaks attention mass onto the phantom zero-vector padding keys instead of concentrating it on the real keys, and that leaked mass is then discarded when the result is trimmed back to S positions. This happens for every sequence length that is not an exact multiple of chunk_size, i.e. essentially always, and it understates the per-key importance score NonCausalAttnPress/CompactorPress use to pick which tokens to evict.

The fix masks with torch.finfo(dots.dtype).min instead, matching the pattern kvzip_press.py already uses for the same kind of key masking, so it stays correct across dtypes (fp32/fp16/bf16).

Fixes #295

Verification

• Added test_non_causal_chunked_attn_does_not_leak_mass_to_padding_keys: a 4-query ragged chunk (2 real, 2 padded) must return per-key scores summing to chunk_size (every query's full softmax unit of mass lands on the 2 real keys). Fails on main (returns ~2.3), passes on this branch.
• pytest tests/presses/ (729 passed, 3 GPU-only skipped) and mypy . / flake8 / black / isort all clean, run in a CPU-only container since this box has no GPU; the Test workflow needs an linux-amd64-gpu-l4-latest-1 runner I could not reproduce locally.
• Did not check whether this changes downstream generation quality on a real model/benchmark, only the masking math itself.
  🤖🤖🤖

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

[SimJeg]

Thanks @AmirF194, @vnchari what do you think ?

[SimJeg]

@vnchari, do you have bandwidth to check this fix on the press you contributed ?