[PyTorch] Normalize FA4 unlimited attention windows

[yashaswikarnati]

Summary

Transformer Engine represents an unlimited attention window with -1. FlashAttention 4 expects None; passing (-1, 0) for causal attention can select an empty local window. Normalize the tuple for non-CP attention and the left/right kwargs for CP p2p, all-gather, and a2a forward and backward calls.

Tests

• Added a unit test for the sentinel conversion, including None and finite windows.
• Added a BSHD/THD causal forward and backward comparison against PyTorch SDPA while forcing FA4.
• AGA 905211 passed the helper unit test against the matching internal TE extension.
• AGA 906233 passed head-256 causal BSHD and THD FA4 b32 forward/backward comparisons with PyTorch SDPA using the cherry-picked TE commit in a full NGC-base image.
• Python compilation, Ruff, git diff --check, DCO, pre-commit, and all reported GitHub CI checks passed. A separate head-128 CP suite case (AGA 906243) was stopped after 45 minutes of compilation without a test result; the target head-256 CP32 training path is running as AGA 906508.

The internal nemotron4 port is one cherry-picked commit on top of 1ac66e5a, proposed in draft MR !4.