[Diffusion][Cosmos3] Enable breakable CUDA graph

[BBuf]

Summary

Enable breakable CUDA graph for Cosmos3 Nano and preserve request warmup for request-only dimensions such as frame count.

The Cosmos-specific prompt padder keeps max_text_seq_len consistent with the padded text mask. Warmup captures the requested text buckets before serving and forwards the current request context through the graph runner.

B300 result

Cosmos3 Nano T2I, 1024px, 35 steps, native backend:

mode	denoise	e2e
eager + BCG	0.880527 s	0.923572 s
torch.compile	0.908647 s	0.952747 s

BCG is 3.09% faster in denoise and 3.06% faster end to end.

Validation

• B300 Cosmos/BCG/server-args: 256 passed, 46 subtests passed
• pre-commit: passed

Performance was measured on commit 9deb6952afa483e38f96385a375b96f463da5303; this PR was rebased and retested on the latest main.

---

CI States

Latest PR Test (Base): ✅ Run #31628635294
Latest PR Test (Extra): 🚫 Run #31630111122

[BBuf]

Closing this candidate after the latest review. The Cosmos3 BCG path improved the measured B300 denoise latency, but the request-warmup lifecycle and model-specific state/test surface are not clean enough to justify the added maintenance risk. I will restore the B300 table to the retained main-branch result and mark this PR as closed.