feat(export): support exporting sub-models from composite models

[xieofxie]

Add --submodel option to winml export for selecting a specific sub-model from composite models (e.g., encoder, decoder).

When a composite model is detected:

• --submodel : exports only that component as {stem}_{name}.onnx
• No --submodel: exports all sub-models automatically
• Non-composite models: unchanged single-model export

Extract _export_single_model() helper shared by both paths and add _resolve_composite_model_components() for registry/auto-detection.

Resolves #1014

[DingmaomaoBJTU]

Took a look — small, focused feature and the tests cover the happy path plus the two error cases. A few things worth addressing before merge: a test-hermeticity concern (the new --submodel docstring example is auto-run and uniquely depends on live Hub resolution), an exception-type consistency nit, and a --submodel + --input-specs dead end. Details inline.

Also, the PR description mentions extracting _export_single_model() and adding _resolve_composite_model_components(), but neither is in the diff — the code uses the existing _run_component_export() / resolve_composite_components(). Worth updating the description so it matches what actually changed.

[DingmaomaoBJTU]

Re-reviewed the fix(export): address PR #1031 review comments commit (3bf05092). All four earlier points are resolved correctly:

1. Test hermeticity — resolve_composite_components is now patched in test_export_help_examples_run (with a .called assertion), so the --submodel example no longer depends on the live Hub. Nice bonus: this makes all the docstring examples hermetic, not just the composite one.
2. Exception type — both --submodel validations now raise click.BadParameter(param_hint="--submodel"), consistent with the rest of the command (usage error + exit code 2).
3. --submodel + --input-specs dead end — the specs are now allowed for a single selected sub-model (input_specs and submodel is None guard), and the multi-component message points at --submodel instead of the --task dead end. I checked test_composite_rejects_input_specs still holds (single-component mock, no --submodel → still rejected, message still contains "composite").
4. Help text — now documents the {stem}_{name}.onnx output.

One follow-up inline (test-strength, non-blocking). Also very minor: the reworded rejection says "exporting multiple sub-models", but the guard fires for any composite without --submodel; it reads slightly off against the single-component mock in test_composite_rejects_input_specs (real composites are all ≥2, so it's only cosmetic). LGTM once the test point is considered.

[DingmaomaoBJTU]

Approving. All four points from the earlier review are addressed correctly and I confirmed there is no regression (test_composite_rejects_input_specs still holds; changes are within the 100-char line limit). The one remaining inline note (strengthening test_submodel_allows_input_specs to assert the specs actually apply) is non-blocking and can be picked up as a follow-up.

Note: I could not execute the suite locally (win_arm64 has no torch wheel), so this is based on code-path analysis; please make sure CI is green before merging.