Add streaming Triton backward and determinism support

[MagellaX]

Summary by cubic

Adds a single-sweep Triton backward for attention (dQ/dK/dV via saved LSE when dropout_p == 0) and makes the Triton forward feature-complete with masks, dropout, deterministic seeding, and ALiBi. This reduces SDPA fallbacks and enables reproducible training/inference behavior.

• New Features
  • Streaming Triton backward when dropout_p == 0; falls back to SDPA if training with dropout.
  • Deterministic mode: set_deterministic(enabled, seed) seeds Philox for reproducible dropout/mask sampling.
  • Forward supports boolean/additive attn_mask, dropout, and ALiBi bias directly in Triton (no SDPA fallback needed).
  • Expanded forward signature: forward(..., dropout_p, alibi_slopes, deterministic); added mask prep for [B, H, M, N].
  • Added tests for mask parity, dropout determinism, and backward parity (with mask + ALiBi); updated docs/README usage.

[cubic-dev-ai]

No issues found across 4 files