[Feature] Migrate AICPU launch to new rtsLaunchCpuKernel interface (BUILD_WITH_NEW_CANN)

[hw-native-sys-bot]

Summary

Migrate the AICPU kernel launch path from the legacy rtAicpuKernelLaunchExWithArgs API to the new rtsLaunchCpuKernel / rtsBinaryLoadFromFile / rtsFuncGetByName interface available in newer CANN versions. This is gated behind BUILD_WITH_NEW_CANN in the pypto codebase and should be adopted in simpler for forward compatibility.

Motivation / Use Case

Current state (simpler):

Both a2a3 and a5 platform backends use the legacy launch path in device_runner.cpp:

// src/a2a3/platform/onboard/host/device_runner.cpp:607
rtAicpuKernelLaunchExWithArgs(
    rtKernelType_t::KERNEL_TYPE_AICPU_KFC, "AST_DYN_AICPU",
    aicpu_num, &rt_args, nullptr, stream, 0);

This requires manually constructing rtAicpuArgsEx_t with kernelNameAddrOffset / soNameAddrOffset, embedding kernel and SO names as fixed-size char arrays in a struct.

New interface (pypto, under BUILD_WITH_NEW_CANN):

pypto has migrated to a cleaner approach using LoadAicpuOp:

1. Load: rtsBinaryLoadFromFile(jsonPath, &optionCfg, &binHandle) — load AICPU op info from a JSON descriptor
2. Resolve: rtsFuncGetByName(binHandle, opName, &funcHandle) — get function handle by name
3. Launch: rtsLaunchCpuKernel(funcHandle, blockDim, stream, &launchCfg, &argInfo) — launch with typed args

Benefits:

• Cleaner API: No manual offset calculations (kernelNameAddrOffset, soNameAddrOffset), no embedded char arrays
• Forward compatibility: The legacy rtAicpuKernelLaunchExWithArgs may be deprecated in future CANN versions
• Consistency: Aligns simpler's host launch path with pypto's approach
• Custom op support: The new interface supports both built-in ops (LaunchBuiltInOp) and custom ops (LaunchCustomOp) through a unified LoadAicpuOp class

Proposed API / Behavior

Add BUILD_WITH_NEW_CANN compile flag support and a LoadAicpuOp-style abstraction:

// New path (when BUILD_WITH_NEW_CANN is defined):
// 1. Generate op info JSON at init time
// 2. Load binary handle: rtsBinaryLoadFromFile(...)
// 3. Resolve function handles: rtsFuncGetByName(...)
// 4. Launch: rtsLaunchCpuKernel(funcHandle, blockDim, stream, &cfg, &args)

// Legacy path (fallback):
// Existing rtAicpuKernelLaunchExWithArgs code unchanged

Scope:

• src/a2a3/platform/onboard/host/device_runner.cpp — AICPU launch in launch_aicpu_kernel()
• src/a5/platform/onboard/host/device_runner.cpp — same pattern
• New header: rts/rts_kernel.h dependency (from CANN toolkit)
• New include dependency gated behind #ifdef BUILD_WITH_NEW_CANN

Reference implementation: pypto/framework/src/machine/runtime/load_aicpu_op.cpp and load_aicpu_op.h

Alternatives Considered

• Keep legacy API only: Works for now, but risks breakage if CANN deprecates the old interface
• Conditional compilation (recommended): Use #ifdef BUILD_WITH_NEW_CANN to support both old and new paths, matching pypto's approach. This allows gradual migration without breaking existing builds

Additional Context

• pypto reference: framework/src/machine/runtime/load_aicpu_op.{h,cpp} and device_runner.cpp
• The new API requires rts/rts_kernel.h header from the CANN toolkit
• AICore launch (rtKernelLaunchWithHandleV2) is unaffected — only AICPU launch changes