feat: add DCU kernel auto optimization task

[fms-zth]

Summary

Add a new dcu-kernel-auto-opt task plugin for automated DCU kernel generation and optimization on Hygon K500SM_AI / gfx928.

The workflow uses multiple isolated agents and GPU workers to generate, benchmark, profile, optimize, synthesize, and validate HIP kernels. The initial production target is INT8 W8A8 GEMM for DeepSeek-V4 TP4/TP8 workloads.

Key changes

• Add the dcu-kernel-auto-opt task plugin, including:
  • task configuration and form schema
  • orchestration pipeline and phase state machine
  • multi-agent worker lifecycle
  • GPU assignment and parallel exploration
  • correctness and performance acceptance gates
  • PMC/ISA analysis support
  • skill and experience storage
  • final candidate synthesis and serial regression validation
• Add the trusted INT8 W8A8 GEMM API contract and HIP variant reference.
• Add the W8A8 benchmark harness and baseline extension template.
• Add WebUI pages for:
  • guided workload/shape configuration
  • worker and iteration status
  • optimization results and reports
  • task cancellation and cleanup
• Integrate the custom shape widget with the next-app form widget registry.
• Improve task process handling:
  • treat zombie processes as stopped
  • expose agent PID and status
  • cancel external agent processes when a task is killed
  • clean up task-owned generated kernel repositories
• Add unit and integration tests for the new workflow.

Workflow

The task follows this high-level pipeline:

1. Parse configuration and validate the immutable operator API contract.
2. Generate or stage the kernel repository.
3. Measure the trusted baseline.
4. Run parallel optimization workers, with one worker assigned to each GPU.
5. Accept candidates only when correctness and performance gates pass.
6. Synthesize accepted worker artifacts into a final implementation.
7. Run serial validation across optimization and fallback shapes.
8. Generate the final report and reusable optimization kno wledge.

Agents are restricted to their assigned source files. Compilation, benchmarking, profiling, artifact verification, and final acceptance remain under control-plane ownership.

Compatibility with next-app

This branch is rebased onto the latest next-app.

The DCU-specific form widget is registered through the new shared form widget registry, and the pipeline has been adapted to the current StateStore.init_or_resume() API.

Validation

Local non-GPU test suite:

200 passed