[Feature] qwen3_32b_decode_mixed.py on A3/A5 platforms with good precision & performance

[zhangqi-chen]

Summary

Track the validation and performance of qwen3_32b_decode_mixed.py on both A3 and A5 platforms.

Motivation / Use Case

1. The mixed-style decode kernel runs correctly on A3
2. The mixed-style decode kernel runs correctly on A5
3. Performance on both platforms is benchmarked and maintained at a competitive level
4. Regressions are caught early as the compiler and runtime evolve

Proposed API / Behavior

• Validate numerical correctness on A3
• Validate numerical correctness on A5
• Ensure performance stays competitive with scope-specific implementations

[zhangqi-chen]

Blockers

• [Feature] Support dual-slot kernel dispatch in Simpler simpler#492
• [Bug] ExpandMixedKernel drops matmul when output is sliced before vector consumption pypto#963
• [Bug] ExpandMixedKernel fails with "Tensor view not found" for V→C pattern in split=UP_DOWN block pypto#965
• [Feature] Support if-guard codegen mode for pl.parallel with dynamic bounds (avoid main+remainder split) pypto#930
• [Pass Bug]SplitVectorKernel fails on singleton broadcast tile loads in split AIV kernels pypto#975

Non-Blockers

• A2A3 TPipe template is incompatible with ptoas 0.24 old-style LocalSlotNum emission pto-isa#67