fix: limit step attention to active cache

[ousamabenyounes]

Summary

• limit step attention to the populated KV-cache prefix
• remove the now-redundant boolean attention mask
• retain contiguous tensors for the documented MPS compatibility path

Fixes #293.

Test verification (RED → GREEN)

The regression test intercepts the SDPA boundary, records the KV sequence
length and mask presence, and still calls the real PyTorch SDPA implementation.

With the new test on upstream/main (production fix absent):

FAILED tests/test_minicpm4_attention.py::test_forward_step_only_attends_to_populated_cache
1 failed, 2 warnings in 2.75s

With the production fix restored:

1 passed, 2 warnings in 2.74s

The production-only revert was repeated after implementation and reproduced
the same failure; restoring the patch returned the test to green.

Full local suite validation:

• baseline suite: 56 passed
• patched suite: 57 passed
• Black: passed for both changed files
• source distribution and wheel build: passed
• changed-line coverage: 100% (2 executable lines, 0 missing)

No local CUDA benchmark was run or claimed.