Refactor: enforce mix strict priority in scheduler dispatch

[poursoul]

Summary

Reshape Phase 4 of SchedulerContext::resolve_and_dispatch from shape-outer/phase-inner into a new dispatch_ready_tasks pass with phase-split semantics and cross-thread idle gating. Applied to both a2a3 and a5 tensormap_and_ringbuffer runtimes.

• MIX strict priority: IDLE-MIX runs first. If mix residual is detected (local_buf + ready_queue), AIC/AIV yield both their IDLE and PENDING stages for this pass; only the next loop iteration (after another completion poll) lets them re-enter.
• MIX-PENDING always considered: gated only on has_idle_in_other_threads(MIX), not on the skip_aic_aiv flag — pending slots keep draining mix even when AIC/AIV are blocked.
• Cross-thread idle gating for PENDING: AIC/AIV-PENDING skip themselves if any peer scheduler thread has an idle core of the same shape, biasing pending fills to threads whose own cores would otherwise sit idle.
• Local-buf flush points: between IDLE and PENDING so peer threads see IDLE-stage release_fanin output; again at function end so PENDING-stage release_fanin output does not carry across iterations.
• PMU single-issue short-circuit and sync_start drain protocol preserved unchanged. a5 picks up the PMU guard alongside the new policy (its prior implementation lacked it).

has_idle_in_other_threads reads peer trackers' core_states_ without explicit synchronization; aarch64 8-byte aligned single-copy atomicity covers the load, and the value is consumed as a scheduling hint (stale reads self-correct on the next iteration).

Testing

• a2a3 sim + onboard and a5 sim + onboard variants build clean (incremental + clang-format + clang-tidy + cpplint)
• cpput a2a3: test_scheduler_state (10/10), test_ready_queue (25/25); a5: test_a5_fatal (3/3) — green
• st on a2a3sim: spmd_multiblock_mix, spmd_sync_start, spmd_sync_start_stress, spmd_sync_start_edge, spmd_sync_start_aiv, spmd_starvation, mixed_example — 7/7 passed (60.96s)
• st on a5sim: same 7 scenarios — 7/7 passed (72.62s)
• Hardware perf bench (benchmark_bgemm) on both arches — recommended before merge

[gemini-code-assist]

Code Review

This pull request refactors the scheduler's dispatch logic to implement a MIX-strict-priority policy, introducing a more structured two-phase dispatch process (IDLE and PENDING) within the new dispatch_ready_tasks method. It also adds cross-thread idle gating via has_idle_in_other_threads. Feedback was provided regarding the implementation of has_idle_in_other_threads, specifically noting that performing cross-thread reads without std::atomic constitutes a data race and undefined behavior under the C++ memory model, regardless of hardware-level atomicity guarantees on specific architectures.