feat(ipc): add DeepSeek4 DSpark shared worker

[davide221]

Summary

• add a deepseek4-dspark-draft backend IPC worker that runs the DSpark proposal graph on a selected HIP GPU
• keep the complete DeepSeek4 target, KV cache, tied head, confidence policy, and sampler in the parent process
• batch all captured target layers into one token-major feature transfer and one acknowledgement
• default DSpark to the existing POSIX shared payload transport, with bidirectional proposal input/output reuse and a stream fallback
• fail initialization when the worker is required but unavailable, and disable incompatible decode-graph reuse while feature capture hooks are active

Stack

#502 ROCmFPX ggml -> #503 DeepSeek4 target -> #496 DSpark runtime -> #505 IPC

The #496 and #505 branches were rebuilt directly on their immediate bases so each pull request contains only its own changes.

Placement

This is stacked on #496. On the tested Strix Halo + R9700 host, the target runs on hip:1 and the DSpark proposal worker runs on hip:0.

The shared transport is host memfd IPC. It is not yet direct hipIpcMemHandle or peer-device memory, and drafting and target verification remain serial.

Validation

• Release HIP build for gfx1151;gfx1201: dflash_server, backend_ipc_daemon, test_server_unit, and test_ds4_dspark_load
• R9700 DSpark load + forward at context 42: 20,480 finite floats, RMS 0.2820
• required shared worker starts on the R9700 and reports transport=shared
• end-to-end greedy requests return coherent output (Paris; 120-token Roman Empire response)
• fix/feat(ds4): ROCmFPX byte-safe decode + fused-decode path + ds4 optimizations #503 fused AR path remains active with the complete stack
• fixed-q3 shared-worker and confidence-adaptive paths both complete coherent generation
• the best recorded shared-worker run remains 23.6 tok/s; throughput varies materially with acceptance rate and device clocks
• IPC mode assertions pass; the full unit binary retains the base branch's unrelated empty-chat-template failure (1 failure across 2,044 assertions)

Follow-ups

• pinned host staging and asynchronous copies
• capability-gated direct HIP IPC / peer-memory handles
• overlap drafting with target verification
• overlap-aware adaptive-width tuning on the reconciled stack