perf(ds4): overlap attention across heterogeneous GPUs

[davide221]

Summary

• adds an opt-in 25% attention-head split for ratio-4 verifier layers
• overlaps the R9700 and Strix branches with a packed destination-stream fork and one deferred residual join
• synchronizes both backends before heterogeneous scheduler teardown
• adds request-windowed rocprof analysis and records the corrected kernel-level findings

Qualification

• same-binary control: 88.538 tok/s
• final two-warmup, seven-run split median: 88.472 tok/s, range 88.309–88.611
• exact expected response hash: 7/7
• measured R9700 overlap: 21.20% to 30.67%

The attention split is throughput-neutral and remains opt-in. Kernel traces show that the peer branch adds more work than the main branch saves, so this PR does not claim a decode speedup. The profiler report also records rejected peer-cache and expert-reuse experiments so they are not repeated.

Validation

• HIP grouped-expert correctness suite passes
• Python profiler tools compile and expose valid CLI help
• git diff --check passes
• restored qualification server checksum matches the proven baseline binary

Stacked on draft PR #577.