Add Reference Tests

Add tests to validate reference implementations in the HuggingFace dataset clone at tmp/flashinfer-trace/. Ground truth is sourced from FlashInfer repository or SGLang when FlashInfer doesn't have the implementation.

Description

This skill creates test cases under tmp/flashinfer-trace/tests/references/ (the HF dataset clone — the in-repo flashinfer_trace/ directory was removed in the trace-dataset refactor) to validate that reference implementations in Definition JSON files produce correct outputs. The ground truth comes from:

1. FlashInfer repository (preferred): Official optimized GPU kernels in tmp/flashinfer/
2. SGLang repository (fallback): When FlashInfer doesn't have the kernel, use tmp/sglang/

Usage

# Test all definitions of a specific op_type
/add-reference-tests --op-type mla_paged
/add-reference-tests --op-type moe
/add-reference-tests --op-type gqa_paged
/add-reference-tests --op-type rmsnorm