Add: validate block_dim against stream resource limit via aclrtGetStreamResLimit

[doraemonmj]

Summary

• Add DeviceRunner::validate_block_dim() (a2a3 + a5 onboard): query the
  stream's CUBE/VECTOR core limits via aclrtGetStreamResLimit and reject a
  block_dim that exceeds hardware capacity, with a clear error showing
  max_block_dim, cube, and vector counts. max_block_dim is derived
  from PLATFORM_AIC_CORES_PER_BLOCKDIM / PLATFORM_AIV_CORES_PER_BLOCKDIM
  (not a hardcoded ratio), and a zero core count is treated as "query
  unavailable".
• When aclrtGetStreamResLimit is unavailable (older firmware) or reports no
  cores, fall back to the static PLATFORM_MAX_BLOCKDIM cap so block_dim
  stays bounded — error stays phrased in block_dim terms.
• Consolidate the onboard block_dim validation (lower bound + capacity
  check) into validate_block_dim(), called from DeviceRunner::run() once
  the device is initialized; removes the old inline checks.
• Drop the block_dim % scheduler_thread_num divisibility check from both
  onboard and sim DeviceRunner::run(). Both runtimes already tolerate an
  uneven split: host_build_graph (AicpuExecutor::assign_cores_to_threads)
  and tensormap_and_ringbuffer (SchedulerContext::assign_cores_to_threads
  / reassign_cores_for_all_threads) assign cores to scheduler threads
  cluster-aligned round-robin and size each thread's tracker from its actual
  cluster count, so block_dim need not divide evenly. Sim keeps the static
  PLATFORM_MAX_BLOCKDIM bound (it has no stream resource query), so onboard
  and sim validation stay consistent.
• Prevents handshake deadlock by failing fast with actionable diagnostics.

Testing

• Sim runtimes (a2a3sim / a5sim libhost_runtime.so) rebuild cleanly
  with the changes; pre-commit (clang-format, clang-tidy, cpplint) passes.
• Hardware: confirm an over-capacity block_dim is rejected with the new
  max_block_dim=... cube=... vector=... diagnostic instead of a
  handshake hang; confirm the PLATFORM_MAX_BLOCKDIM fallback path on
  firmware without aclrtGetStreamResLimit.
• Hardware: a non-divisible block_dim (e.g. block_dim=5,
  aicpu_thread_num=4) runs to completion (sanity-check the round-robin
  core split now that the divisibility gate is gone).

Review notes addressed

• Magic /2 AIC:AIV ratio → platform constants.
• Lost PLATFORM_MAX_BLOCKDIM fallback restored (was only a LOG_WARN).
• Zero core count from a "successful" query no longer reports max_block_dim=0.
• Divisibility check removed from sim too, so onboard/sim no longer diverge.
• validate_block_dim doc comment updated to match the actual contract.

Notes

• PR-title typo "GetSteamResLimit" fixed.
• Runtime::get_orch_built_on_host() is kept — this PR only drops its one
  caller in the platform layer (the divisibility check). The flag is still
  load-bearing inside tensormap_and_ringbuffer (aicpu_executor.cpp,
  scheduler_cold_path.cpp orchestrator_done_, runtime_maker.cpp
set_orch_built_on_host) and host_build_graph (runtime.h returns
  true); removing it is out of scope and would change runtime behavior.

[gemini-code-assist]

Code Review

This pull request refactors block_dim validation into a dedicated validate_block_dim method across DeviceRunner implementations, introducing dynamic resource limit checks for Cube and Vector cores via aclrtGetStreamResLimit. Review feedback points out that critical divisibility checks for scheduler threads were lost during the consolidation, which could lead to handshake deadlocks or logic errors. Additionally, the reviewer noted that error logs may be misleading by reporting zero available cores if the hardware resource query fails, suggesting more conditional logging.