[qwen] Drop fp32 LayerNorm casts in Qwen3.5 text model on ORT >= 1.26 - #361

Ports microsoft/onnxruntime-genai#2101: Qwen35TextModel was forcing every RMSNorm to cast IO to fp32, inserting ~216 Cast nodes (108 to-fp32 + 108 to-fp16) around SkipSimplifiedLayerNormalization in a 24-layer build. Upstream measurements show +13% gen tps on RTX 5080 and +9% on Intel iGPU with output remaining structurally identical, so the casts are removed — but only when the runtime is recent enough.

Changes

• modelbuilder/builders/qwen.py — gate the five layernorm_attrs["cast"] assignments (use_fp32, root_input, skip_input, output_0, output_3) in Qwen35TextModel.__init__ on the installed onnxruntime version. On ORT >= 1.26 (nightly) the assignments are dropped, matching upstream onnxruntime-genai#2101. On older ORT releases the explicit fp32 cast wrapping is preserved, since the native fp16 SkipSimplifiedLayerNormalization kernel loses precision across Qwen3.5's 36+ layers.
• tests/fast/test_random_qwen3_5.py — add test_qwen3_5_fp16_no_layernorm_fp32_casts: builds a tiny fp16 Qwen3.5 ONNX model and walks the graph to assert no Cast-to-fp32 feeds, and no Cast-from-fp32 consumes, any SkipSimplifiedLayerNormalization op. Verified to fail (15 offending casts) when the removed assignments are reintroduced. The test is skipped when running on onnxruntime < 1.26.

Diff sketch

# Qwen35TextModel.__init__
self.layernorm_attrs["add_offset"] = 1

# Keep RMSNorm IO in the model's native dtype on ORT >= 1.26 (see
# onnxruntime-genai#2101); older ORT still needs the fp32 cast wrapping.
if _ort_version() < (1, 26):
    self.layernorm_attrs["cast"]["use_fp32"] = True
    self.layernorm_attrs["cast"]["root_input"] = True
    self.layernorm_attrs["cast"]["skip_input"] = True
    self.layernorm_attrs["cast"]["output_0"] = True
    self.layernorm_attrs["cast"]["output_3"] = True