feat(moe): add a generic heterogeneous stage planner

[davide221]

Summary

• keep the generic, model-neutral two-owner stage planner and owner-local residual primitive
• add the complete R9700 + Strix Halo ROCm 7.2.4 profile: routing inputs, hipBLASLt table, checksum guards, and a one-command qualifier
• fix a rebase regression that split heterogeneous 2K prefill into 1,920 + 128 tokens, changed the sparse-prefill state, and ended generation after two tokens
• keep the repaired path as one wide batch while preserving required snapshot-boundary splits
• report authoritative model-side DSpark decode separately from the historical client transport diagnostic

Reproduce

Build for both GPU architectures, then run with only the two model paths:

cmake -S server -B server/build-hip-dual \
  -DCMAKE_BUILD_TYPE=Release \
  -DDFLASH27B_GPU_BACKEND=hip \
  -DDFLASH27B_HIP_ARCHITECTURES='gfx1151;gfx1201' \
  -DDFLASH27B_TESTS=ON

cmake --build server/build-hip-dual -j \
  --target dflash_server test_tokenizer_harness test_deepseek4_unit

TARGET_MODEL=/path/to/target.gguf \
DRAFT_MODEL=/path/to/dspark-draft.gguf \
BUILD_DIR="$PWD/server/build-hip-dual" \
server/scripts/qualify_ds4_q5_r9700_strix.sh

The wrapper defaults to 2 warmups + 7 measured 2K/128-token requests and fails on a short or non-exact response. See server/docs/DS4_R9700_STRIX_PROFILE.md.

Validation

• head: 682e419cc21c9de4585e705b2bb21f7dc3274efa
• dual-architecture HIP Release build: passed
• full test_deepseek4_unit: passed, including GPU oracles and the new capture-policy cases
• test_server_unit heterogeneous_stage: 5 passed, 0 failed
• server-summary Python tests: 2 passed
• exact model qualification: 7/7 measured responses produced 128 tokens and SHA-256 0f785a7ffa406498aafb14553966eaed0f52220fed0f7cc016b66921d104d194
• authoritative server decode: 73.7 tok/s median, 73.5-73.8 range, acceptance 1.00
• historical client diagnostic: 83.403 tok/s median, 83.122-83.509 range

Review finding about the prior 89 tok/s result

The earlier 89.876 value was not model-side throughput. Its server median was 73.2 tok/s. The client timer began at the first non-empty streamed text event but counted all 128 completion tokens; a slow first rejected speculative step fell outside the timed interval and inflated the result.

The repaired current path has full first-block acceptance and is slightly faster model-side (73.7 versus 73.2), although the legacy client number is lower. The qualifier now writes server-decode-summary.json so future comparisons use the correct metric.

Measured-slower attention and shared-width splits remain disabled.

Stacked on #575.