CI: add numerical parity coverage for the static-cache export path

[titaiwangms]

Gap

PR #328 adds the static-cache ONNX Attention export path (is_causal=0 + a phase-split If). Its correctness is covered in CI on two axes, but a third is missing:

• Structural correctness — CI-gated by tests/build_graph_test.py (CPU, runs per-PR).
• Runtime-CUDA kernel routing — now gated by tests/static_cache_decode_test.py in the integration-fast GPU job (added in Fix Phi-3.5 ONNX Attention (static-cache) + GQA fp16 export bugs #328).
• Numerical parity — not covered. The committed tests/synthetic_parity_test.py only exercises the dynamic path (is_causal=1). The 20/20 static-path parity validated during Fix Phi-3.5 ONNX Attention (static-cache) + GQA fp16 export bugs #328 development was a manual QA run (parity_llama.json), not a committed test.

Proposed follow-up

Author a synthetic static-cache parity test:

• Build a tiny in-memory synthetic model (same approach static_cache_decode_test.py uses).
• Export the static-cache variant and a dynamic reference.
• Assert token-id and/or logit parity within tolerance over N decode steps.
• Wire it into a GPU CI tier (static-cache decode requires CUDA).

This complements the runtime-kernel routing test added in #328 — that test checks the right kernel runs; this one would check the numbers are right.

Refs: #328