fix(rocm): make CPU/Hybrid MoE graph replay safe

[zihaomu]

Summary

Fixes #350, where CPU/Hybrid MoE produces silently incorrect output during ROCm graph replay while eager execution remains correct.

• fail closed on ROCm: if the native stream-memory handshake does not pass a real capture/instantiate/replay probe, disable graph capture and continue on the correct eager path
• add the ROCm 7.14 native handshake with hipMallocSignalMemory and explicit graph batch-memory-op nodes
• keep HIP signal and graph-node parameter storage owned by CpuMoeExecutor, so their lifetime covers every graph replay without module-global growth
• size per-layer flag capacity from the actual configured graph batch sizes, including configurations with more than 16 sizes
• preserve the established CUDA module-level memop_submit/memop_sync path to limit NVIDIA regression risk
• document the CUDA/ROCm synchronization difference and the ROCm fallback behavior

Dependency and scope

Depends on #132. This branch is based directly on the current #132 head (c0713e4). #132 itself is unchanged.

This is intentionally a Draft stacked PR while the native ROCm graph path receives RDNA3 and long-running model-serving stability validation. Until #132 merges, GitHub will also show the dependency commits in this PR; the diff will collapse to this single follow-up commit after the base lands.

CUDA and ROCm retain the same externally visible CPU-MoE ordering and results. Only their GPU/CPU synchronization implementations differ:

Backend	Handshake implementation
CUDA	Existing mapped-pinned per-slot arrays through module-level cuStreamWriteValue64/cuStreamWaitValue64 wrappers
ROCm 7.14	Executor-owned HIP signal memory plus explicit graph batch-memory-op nodes

Validation

• PyTorch 2.11.0 + ROCm 7.14 on one Radeon AI PRO R9700: targeted safety and replay suite, 21 passed
• capture/replay exercised across 19 graph batch sizes, covering the former fixed-16-slot boundary
• eager plus repeated graph replay covered bf16, MXFP4, DS-FP4, and GGUF Q4_0 CPU-MoE formats
• ROCm 7.2.4 / PyTorch 2.10 compatibility build passed and remains fail-closed where graph batch-memory-op APIs are unavailable
• CUDA module API and dispatch routing are covered by unit tests; NVIDIA hardware validation remains required before this Draft is marked ready
• git diff --check passes

Draft exit criteria

• reproduce the replay suite on RDNA3
• complete repeated graph rebuild/replay and long-running serving stability checks
• obtain NVIDIA CI coverage for the preserved CUDA path