Clean up native attention nits (cc threshold comment, duplicate reason, stale docstrings)

[adohe]

Items

• FLASH_ATTENTION_MIN_CUDA_CAPABILITY = (8, 0) disables flash on all Turing (sm_75/T4) though flash-attn 2.x forward supports some head dims on Turing — add a comment explaining why 8.0 not 7.5.
• flash_attention_unsupported_model_reason can append "qk head dim ..." twice when a model sets both head_dim and qk_nope_head_dim/qk_rope_head_dim — dedupe the reason.
• common.py top docstring still mentions converting between "FlashAttention and SDPA" conventions, but sdpa_window_size was removed — clean up stale text.

Acceptance

• Comments/docstrings reflect current behavior; reason list deduped. No behavior change.