Fix: guard idle-dispatch do-while against empty cluster mask

[chenshengxin2026]

Summary

Fixes #565 — out-of-bounds core_id_map_[-1] access in the idle-dispatch batch loop.

• When pop_ready_tasks_batch returns multiple tasks (got > 1) and the first task's do-while drains all idle clusters (common when logical_block_num >> cluster_count, e.g. paged-attention with block_num=256 on 24 clusters), subsequent tasks enter the do-while unconditionally — because do { ... } while(...) executes the body before checking the guard — calling pop_first() on an empty bitmask, which returns -1
• -1 is used as cluster_offset → core_id_map_[-1] (OOB read) and 1ULL << -1 (undefined shift), corrupting core_states_ and stalling the scheduler
• Fix: add has_value() guard before the do-while; when clusters are exhausted mid-batch, re-enqueue the current and remaining tasks and break. Zero overhead on the common path (single predicted-taken branch)
• Applied to both a2a3 and a5 variants

Test plan

• New test spmd_batch_dispatch_oob: submits 2 back-to-back MIX tasks with block_num=48 (2x cluster count) to force got=2 and trigger the OOB scenario — passes with fix, would crash/stall without
• Existing spmd_multiblock_mix test: passes (no regression)
• Hardware device test on a2a3

[gemini-code-assist]

Code Review

This pull request addresses a batch dispatch out-of-bounds (OOB) issue in the aicpu_executor by adding a guard to check for available clusters before popping from valid_cluster_states. If clusters are exhausted by a preceding task in a batch, the remaining tasks are now correctly re-enqueued. The PR also includes a comprehensive regression test suite consisting of AIC/AIV kernels, orchestration logic, and a Python test script. Feedback suggests optimizing the re-enqueueing logic by using push_batch for bulk operations instead of individual push calls to improve efficiency and reduce queue contention.