Add: parallel SPMD block dispatch across schedulers

[zhusy54]

Summary

• Collapse the serial SPMD dispatch staircase: dispatch_shape claims a
  contiguous block range, pushes the slot back to the ready queue early,
  then performs the expensive dispatch_block calls so other schedulers
  concurrently claim the remaining blocks of the same SPMD task
• Thread block_idx explicitly through build_payload /
  dispatch_subtask_to_core / dispatch_mix_block_to_cluster /
  dispatch_block to avoid the race once the slot is pushed back early
• Apply the same optimization to both a2a3 and a5 schedulers (a5 adapts
  to its Runtime* first-arg and s_block_idx/s_block_num fields)
• Add spmd_parallel_dispatch_demo (a2a3) for L2-swimlane comparison

Testing

• a2a3sim numerical PASS; a2a3 hardware swimlane PASS — each task's 24
  blocks split 8/8/8 across the 3 schedulers with overlapping
  first-dispatch times
• a5sim numerical PASS: spmd_multiblock_mix, spmd_sync_start_stress,
  spmd_starvation

Swimlane comparison

Before：

After：

Related Issues

Closes #818

[gemini-code-assist]

Code Review

This pull request optimizes SPMD parallel dispatch in the a2a3 and a5 runtimes by allowing scheduler threads to claim block ranges and return tasks to the ready queue immediately, enabling concurrent execution across threads. It includes a new QK matmul kernel, orchestration logic, and a demo test case. Feedback recommends using more robust types for size calculations to avoid potential truncation and simplifying std::min calls by removing redundant template arguments.