Add: a5 chevron launch path for AICore SIMT validation

[ChaoZheng109]

Summary

Two commits that together unblock SIMT execution on a5:

1. Fix: set AICore localMemorySize for a5 SIMT launch — define PLATFORM_AICORE_LOCAL_MEMORY_SIZE = 229376 (224 KB) and pass it through rtTaskCfgInfo_t::localMemorySize in the existing rtKernelLaunchWithHandleV2 path.
2. Add: a5 chevron launch path for AICore SIMT validation — parallel launch path using the bisheng chevron syntax (kernel<<<numBlocks, dynamic_shmem_sz, stream>>>). The kernel body delegates to the same aicore_execute() handshake loop so AICPU keeps dispatching tasks via registers. Enabled with SIMPLER_USE_CHEVRON_LAUNCH=1; default behavior unchanged.

The chevron form is required for SIMT because it lowers to LaunchAscendKernel, which programs the AICore local memory window through the SIMT-specific runtime path.

aicore_execute moved from .cpp to header

The chevron mix kernel needs aicore_execute visible on the same TU it is called from — ld.lld (inside bisheng) cannot resolve .cube/.vector references across separately compiled TUs, and bisheng -c -o file.o rejects multiple input sources. So:

• src/a5/runtime/{tensormap_and_ringbuffer,host_build_graph}/aicore/aicore_executor.cpp → renamed to aicore_executor.h, signature changed to inline __aicore__.
• The legacy kernel.cpp and the new chevron_launch.cpp both #include "aicore_executor.h" and emit their own instantiation. The host SO uses one launch path at a time, so the device-side duplication is benign.
• build_config.py exposes each runtime's aicore/ directory to the aicore and host targets so both compilers find the header. No Python build-driver plumbing needed.

Testing

• Build on a5 hardware (CANN 9.x): pip install --no-build-isolation . produces chevron_launch.o and host_runtime.so
• nm libhost_runtime.so | grep launch_aicore_chevron shows the symbol
• Existing a5 hardware test still passes without env var
• Same test with SIMPLER_USE_CHEVRON_LAUNCH=1 exercises the chevron path; log shows using chevron (<<<>>>) launch path

[gemini-code-assist]

Code Review

This pull request introduces a new constant PLATFORM_AICORE_LOCAL_MEMORY_SIZE (224 KB) in platform_config.h and updates the DeviceRunner::launch_aicore_kernel function in device_runner.cpp to utilize this constant when configuring the AICore task's local memory size. I have no feedback to provide.