feat(compile): add CGIR to CG Output IR compiler mode

[fangyangci]

Summary

Add a CGIR to device-targeted CG Output IR compiler mode to winml compile, while preserving the existing ONNX-to-EPContext path.

winml compile `
  -m input.mlir `
  --compiler cgc `
  --device gpu `
  -o output.mlir

--compiler cgc is optional for a single .mlir or .mlirbc input; the CLI infers the CGC compiler from those suffixes. It remains available to select the compiler explicitly or for CGIR files with another suffix.

• Route persistent CGIR compilation through the session-backed WinMLRuntimeSession.compile_to_file() API.
• Share Runtime compiler/options ownership with the existing ONNX-to-CGC session path.
• Select a GPU through DXCore, with explicit --device-luid binding when needed, without requiring an ORT execution provider.
• Call the Runtime target's ModelCompiler.compile_to_file() with CompiledModelForm.DEVICE_TARGETED.
• Reject incompatible EP/config/embedding options instead of ignoring them.
• Publish the compiled artifact, Runtime resource companions, validated I/O metadata, and a hash-bound compile receipt with overwrite protection and rollback.
• Remove stale destination metadata when an overwrite no longer produces metadata, while preserving rollback safety.
• Update compile command documentation and CGC onboarding examples.

Output and validation semantics

The output is MLIR bytecode, including when the filename uses .mlir; .mlirbc is also accepted. This change does not add a text-dump option.

Default validation reloads the output, builds a Runtime pipeline, and compares I/O schemas through the existing stage-schema compatibility helper. Dynamic dimensions reported as either 0 or UINT64_MAX are normalized before comparison. Validation does not run inference, certify numerical accuracy, or execute a benchmark. --no-validate skips that reload/build check, not compilation.

The new mode currently supports GPU only. Existing ort, ort_session, and qairt compilation behavior remains unchanged. A compatible preview Windows ML Runtime is required.

Verification

• 188 directly related unit tests passed across compile CLI/compiler, Runtime session, CGC export, and Foundry coverage.
• mypy -p winml.modelkit passed for all 464 source files.
• Changed-file Ruff checks passed; GitHub lint, test, and CodeQL checks passed.
• The adjacent commands/compiler/session/CGC-export suites completed with 3108 passed and 7 skipped; two unrelated local hardware/environment tests failed because OpenVINO was already registered and NPU PDH counters were unavailable.
• 3 native GPU end-to-end cases previously passed: generated text CGIR, generated bytecode CGIR, and a local FP32 ASG model's exported CGIR.
• Native verification used Python 3.11 x64, windowsml 2.8.75.dev0, onnxruntime-windowsml 1.30.0.202609102321, and NVIDIA GeForce RTX 5070.

No model weights, benchmark reports/results, or local runtime packages are included in this PR.

[Qiong Wu (qiowu) (DingmaomaoBJTU)]

I found two validation-path regressions that can reject otherwise valid compiled artifacts; details are inline.

[Zhenchao Ni (zhenchaoni)]

Requesting changes primarily because CGC compilation needs to use WinMLRuntimeSession, consistent with the existing WinMLSession / WinMLQairtSession convention. The three inline comments cover that architectural requirement, stale metadata on overwrite, and the compiler type-narrowing failure.

ONNX input support for --compiler cgc is not a requirement for this PR and is not a blocker.

Please also keep this change incremental: reconsider whether the additional Foundry pre-parse, text/regex classification, and hash receipt are necessary for exposing Runtime compilation. Preserve necessary resource-companion handling and overwrite safety, but avoid a parallel Runtime compilation implementation. The exporter metadata format change from "cgc-input-ir" to "cgir" appears unrelated to this feature and should preferably be left out unless a direct dependency requires it.