[WebGPU] Enable graph capture for Gemma4-class decoder models

[feich-ms]

Note: This PR is intentionally kept as draft and will NOT be merged.
The GqaMaskReformattingToGraphCapture transformer is a temporary workaround for Gemma4 models that were built without enable_webgpu_graph=true (the genai builder does not yet have a Gemma4 dispatch entry). Once the genai team adds Gemma4 builder support and official models are converted with graph capture enabled, this ORT-side transformer becomes unnecessary. Until then, this branch serves as a local development tool for WebGPU graph capture on Gemma4.

---

Graph capture (GC) requires every kernel and buffer interaction in the captured replay to be deterministic and self-contained. Several existing WebGPU paths violated that — they relied on CPU-side branching or opset-open registration. This change makes the EP GC-clean for Gemma-class decoders without regressing the non-GC path.

Five coordinated changes, each gated so non-GC sessions are unaffected:

1. ConstantOfShape kernel + opset-by-opset registration New onnxruntime/core/providers/webgpu/generator/constant_of_shape.{cc,h}. Uses bitcast<output_value_t>(uniforms.value) so a single shader covers all dtypes. Shape input marked OrtMemTypeCPU. Registered explicitly at versions 9, 20, 21, 23, 24 because KernelRegistry::VerifyVersion treats open-ended SinceVersion(N) as exact-match against node.SinceVersion().

2. GqaMaskReformattingToGraphCapture transformer (gated) — shrunk from 433 to 231 lines: simplified matching logic (inlined chain matching with early-continue, replaced TryResolveConstantInt with compact IsConstantInt, merged initializer helpers). Rewrites the int64 GQA-mask construction subgraph into a GC-replayable form (Cast int64 to int32 + ReduceSum/Sub/ReduceMax). Registered in graph_transformer_utils.cc only when ep.webgpuexecutionprovider.enableGraphCapture == 1, so default sessions see no transform.

3. Where int64 scalar path tensor/where.{cc,h}: added is_scalar_ shader variant that reads the condition word and extracts the selected bit, then reads scalar a/b. Drives component=1 output for int64 (stored as vec2). Type constraints widened to include int64_t. CacheHint disambiguates broadcast x int64 variants.

4. Equal / binary-elementwise int64 path + widened type constraints math/binary_elementwise_ops.{cc,h}: int64 shader path packs four scalar reads per global_idx into a vec4 to keep output component=4 while inputs are component=1 (int64 cannot vector-load). Equal registered for int64 across versions 7-10, 11-12, 13-18, 19+.

5. PrepareIndirectDispatch helper + widened use_indirect_dispatch contrib_ops/webgpu/bert/flash_attention.{cc,h}: new PrepareIndirectDispatchProgram single-thread shader writes [tile_count, num_heads, 1] from seqlen_k[0]+1. use_indirect_dispatch is now true when GC is on OR the GPU-resident seqlens_k sentinel is 0, so the dispatch shape no longer leaks through the CPU. Added an explicit PrepareIndirectDispatch call on the kv_empty path because CopyKVCache is skipped there under GC.

Note: GraphCacheManager buffer quarantine (previously listed as change #5) was verified as NOT load-bearing — graph capture produces correct output with PR #27453's immediate buffer reuse. The quarantine has been reverted.

Verified on Gemma4 E2B Q4_K_M (opset 24, com.microsoft 1, 20 KV-shared layers): GC=ON produces byte-identical output to GC=OFF and improves single-batch decode throughput substantially. Non-GC sessions show no behavioral change.

[github-actions]

You can commit the suggested changes from lintrunner.

[github-actions]

Suggested change

	int data_type, const void* data, size_t byte_len) {
	int data_type, const void* data, size_t byte_len) {