Align pattern support with node-level checks (scoped 15-pattern rules)

[fangyangci]

Artifact pattern generate:https://github.com/gim-home/ModelKitArtifacts/pull/156

Summary

• Enable rules checks for the scoped 15 patterns (alternative target patterns + check patterns).
• For pattern cases with node count > 1, run node-level precheck first; if all nodes are supported, directly use the supported result. To keep pattern check results consistent with split node-level check results on supported cases.
• Unify and stabilize check_result output via centralized sanitization and backward-compatible import path.

Scope

• available_ops_all.json has 19 pattern definitions in total.
• This PR intentionally enables rules checks for the scoped 15 patterns, not all 19.

Consistency Note

• Helper path without node-level precheck exists in:
  • ModelKitArtifacts/scripts/check_ops_run_helps/debug_case_helper/debug_pattern_help.bat
• Results between the helper path and the node-level precheck path have been verified as consistent.

Validation

• uv run ruff check src/winml/modelkit/analyze src/winml/modelkit/pattern tests/unit/analyze
• uv run mypy -p winml.modelkit
• uv run pytest tests/unit/analyze/runtime_checker/test_ep_checker.py tests/unit/analyze/test_check_ops.py -q

[DingmaomaoBJTU]

Reviewed the logic statically (didn't run the suite). Overall the rules-first prefilter plus the sys_info / yield_skipped cleanups look coherent — I verified no remaining callers pass the removed sys_info/yield_skipped args, device_type.name is the right value to feed RuntimeCheckerQuery, and the synthetic check_result shape matches the real {compile, run} one. Left inline notes; the main one is a design question about skipping the real EP run purely on node-level rules.

Two nits not worth their own threads: available_ops_all.json gained two trailing blank lines at EOF, and _RulesPrefilterProtocol's method body reads better as ... than pass.