XNNPACK graph runtime: in-tree operator interface

[GregoryComer]

Adds the Operator base interface and the create_operator factory. The factory
is intentionally stubbed: no in-tree operators are wired up yet, so the graph
runtime supports only XNNPACK-delegated subgraphs. The concrete operators and
their kernels land in a later change.

Authored with Claude.

[GregoryComer]

Stack from ghstack (oldest at bottom):

• XNNPACK graph runtime: backend integration behind use_graph_runtime #20256
• XNNPACK graph runtime: flatbuffer graph deserialization #20255
• XNNPACK graph runtime: executor engine #20254
• XNNPACK graph runtime: memory planning #20253
• XNNPACK graph runtime: execution planning #20252
• XNNPACK graph runtime: partitioner #20251
• XNNPACK graph runtime: arena allocator and shape environment #20250
• -> XNNPACK graph runtime: in-tree operator interface #20249
• XNNPACK graph runtime: graph IR and builder #20248
• XNNPACK graph runtime: core tensor and quantization types #20247

[pytorch-bot]

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/20249

• 📄 Preview Python docs built from this PR

Note: Links to docs will display an error until the docs builds have been completed.

❗ 1 Active SEVs

There are 1 currently active SEVs. If your PR is affected, please view them below:

• [ROCm] MI350 CI jobs will have longer queue times due to CI migration

This comment was automatically generated by Dr. CI and updates every 15 minutes.