Consolidate determinism into configure_determinism and clarify fused AdamW behavior

[Sam-v6]

Motivation

• Centralize seeding and determinism knobs so reproducible runs and profiling-driven deterministic runs use a single code path.
• Ensure the same seeded torch.Generator is produced and used for deterministic DataLoader shuffling.
• Allow strict determinism only when requested (e.g., during profiling) to avoid unnecessary throughput loss in normal runs.
• Make the reason to disable fused optimizer kernels in deterministic mode explicit.

Description

• Added configure_determinism(deterministic: bool, seed: int = 42) which sets PYTHONHASHSEED, Python/NumPy/PyTorch seeds, creates and returns a seeded torch.Generator, and toggles cuDNN/TF32/deterministic algorithm flags based on deterministic.
• Replaced the old seed setup in run_trial with g = configure_determinism(deterministic, seed=42) where deterministic is driven by args.profile.
• Passed the returned generator g into DataLoader(..., generator=g) and set the optimizer to fused=not deterministic for AdamW, with an updated comment explaining the non-deterministic reductions.
• Clarified inline comments around fused AdamW determinism to address earlier reviewer questioning.

Testing

• No automated tests were executed for these changes.
• No CI runs were invoked as part of this PR.

---

Codex Task