Update: auto-resolve CallConfig.block_dim to max stream capacity

[ChaoWao]

Summary

• CallConfig::block_dim default flips from 24 to 0 (sentinel for "auto"); DeviceRunner::run resolves 0 at launch time to the max the AICore stream can host.
• Onboard runners query aclrtGetStreamResLimit(CUBE_CORE / VECTOR_CORE) and pick min(cube/AIC_per_bd, vector/AIV_per_bd), falling back to PLATFORM_MAX_BLOCKDIM if the ACL call fails. Sim runners short-circuit to PLATFORM_MAX_BLOCKDIM since there is no per-stream resource query.
• scene_test.py's implicit block_dim fallback (1) is dropped in favour of 0 so cases that omit block_dim exercise the new auto path.

Behavior change

• Any caller constructing CallConfig() without setting block_dim previously got 24. They now get the stream's max — 36 on a full-die a5 stream, 24 on a2a3, possibly smaller on resource-limited streams.
• validate_block_dim now applies the PLATFORM_MAX_BLOCKDIM clamp on both the ACL-success path and the ACL-fallback path (previously only the fallback was clamped). This is a defensive narrowing — the runtime handshake/scheduler arrays are statically sized to RUNTIME_MAX_WORKER = PLATFORM_MAX_BLOCKDIM * PLATFORM_CORES_PER_BLOCKDIM, so accepting an ACL-reported value larger than the platform cap would over-run those static arrays. No observed-hardware impact today, but explicit values above the platform cap now error early instead of corrupting later.

Why

Most callers want "use everything the stream allows" and were hand-picking 24 arbitrarily. Centralizing the policy in DeviceRunner removes the magic number from user code and adapts automatically when stream capacity is constrained (CPU partitioning, model-shared streams, etc.).

Test plan

• Local sim build (pip install --no-build-isolation -e .) passes
• tests/ut/py/test_chip_worker.py::TestCallConfig updated for new default; pre-commit hooks (clang-tidy / cpplint / pyright) clean
• spmd_basic ST gains a Case2_AutoBlockDim on both a5 and a2a3 that omits block_dim, so onboard CI exercises query_max_block_dim end-to-end
• simpler_setup/scene_test.py fallback flipped from 1 to 0 so future cases that omit block_dim also exercise the auto path

[gemini-code-assist]

Code Review

This pull request introduces an "auto" resolution feature for block_dim by changing its default value from 24 to 0. When set to 0, the DeviceRunner dynamically determines the maximum allowable block_dim based on stream resource limits or platform constants. Feedback from the reviewer highlights a loss of diagnostic logging information regarding specific core limits during validation and identifies performance redundancies where resource limits are queried multiple times during the launch path.