[LZ-QWQ]

Related: #697

ROCm disk PLE currently uses gate fallback. This PR adds HIP wait-sync and prevents deadlocks when submitting large graphs with wait-sync.

It adds HIP signal memory, stream memops and capture probing (auto probes; wait requires support). Use HIP signal memory for Engram's wait-sync flag.

#697 uses early cuGraphUpload to avoid the first-launch deadlock on CUDA. In our ROCm 10.1.0 large-graph checks, hipGraphUpload followed by stream synchronization did not prevent the deadlock. This may be related to the current HIP upload implementation: ihipGraphUpload returns after handle validation and stream resolution, while hipGraphUpload still marks preparatory work as TODO. To avoid this deadlock, this PR starts a HIP-only CPU worker before replay to fill PLE rows and signal the flag independently of graph submission.

CUDA's existing allocation, execution and synchronization logic is unchanged.

Validation

The baseline uses main's unchanged production code, with the new 16,384-kernel test included in the test suite.

Test files:

• tests/kernels/test_row_store.py
• tests/models/qwen4_exp/test_ple_disk.py
• tests/models/deepseek_v41/test_engram.py

GPU	Runtime	Main + test	PR
RTX4090	CUDA 13.0	18 passed	18 passed
W7900D	ROCm 10.1.0	15 passed, 3 skipped	18 passed, 0 skipped
R9700_x8	ROCm 10.1.0	15 passed, 3 skipped	18 passed, 0 skipped

The new regression and two existing Engram worker cases now run on ROCm.

Performance reference

Qwen/Qwen3.8-Flash-Next-FP8, 256 input / 128 output tokens, single GPU/request, MoE offload cache 1024, graph batch 1, ROCm 10.1.0. Compare main/gate, PR/gate and PR/wait under identical settings.

GPU	Metric	Main/gate	PR/gate	PR/wait	Wait vs PR gate
R9700_x8	Decode tok/s	17.190	17.172	19.365	+12.77%
R9700_x8	Request (s)	10.497	10.465	9.562	-8.63%
W7900D	Decode tok/s	12.676	12.665	12.780	+0.91%
W7900D	Request (s)	14.529	14.542	14.445	-0.67%

Each configuration was tested in two independent runs, with three measurements after warmup per run. The table reports the average of the two run medians.