[Feature] Improve async DCP checkpoint reliability and early failure …

[VincentCheungKokomo]

[Feature] Improve async DCP checkpoint reliability and early failure detection

• Add warmup_async_save_dcp() to pre-initialize checkpoint daemon before training, surfacing EADDRINUSE errors immediately instead of mid-training
• Add _check_async_save_health() for per-step non-blocking failure detection on both async DCP
• Coalesce per-tensor shared memory into per-dtype buffers to reduce fd count from ~3000 to ~2, fixing "received 0 items of ancdata" errors
• Remove retry logic in commit_async_save() (warmup makes it unnecessary)
• Allow snapshot saves to use the async checkpoint path