feat: Add dynamic axes export support

[xieofxie]

Summary

• Add a first-class --dynamic-axes option for winml export.
• Allow symbolic string dimensions in --input-specs and infer matching dynamic ONNX input axes.
• Normalize/validate dynamic axes mappings, including JSON string axis keys.
• Keep static shapes as the QNN-safe default while making requested dynamic batch informational instead of warning-shaped.
• Document dynamic axes, symbolic input specs, and QNN compatibility caveats.

Fixes #929

Validation

• uv run ruff check --fix src\winml\modelkit\onnx\io.py src\winml\modelkit\export\config.py src\winml\modelkit\commands\export.py src\winml\modelkit\export\htp\exporter.py tests\unit\export\test_config_validation.py tests\unit\export\test_pytorch_export.py tests\unit\commands\test_export.py
• uv run pytest tests\unit\export\test_config_validation.py tests\unit\export\test_pytorch_export.py tests\unit\commands\test_export.py

[DingmaomaoBJTU]

Nice, well-scoped addition — the docs, config validation, and the end-to-end symbolic-shape export test all look solid, and the layered --dynamic-axes > --export-config precedence is clear.

A few consistency nits below around the dynamic-batch messaging and symbolic-dim validation. Nothing blocking.

[DingmaomaoBJTU]

Took a pass over the dynamic axes support. The normalization / merge / inference split in export/config.py is clean, symbolic dims resolving to size 1 for the dummy tensor is a sensible default, and the test coverage is solid. A few non-blocking notes inline: an error-wrapping inconsistency in the --export-config path, a cross-module private-method call, and a small question about the conflict-on-merge UX. Nothing blocking from my side.

[DingmaomaoBJTU]

Did a full holistic pass over the complete change set (not just the incremental delta) and everything checks out. Approving. ✅

What I verified this round:

• All three of my previous comments were addressed cleanly — the --export-config error-handling is now consistent with --shape-config/--dynamic-axes, concrete_shape() was promoted to a public method used by both call sites, and the conflict-on-merge behavior is intentional with a clear CLI error message.
• Symbolic (str) dims don't leak into numeric code paths: the batch_size comparison in __post_init__ guards with isinstance(spec.shape[0], int), dummy generation goes through concrete_shape(), and the metadata embedding (to_dict() + json.dumps) round-trips symbolic shapes correctly.
• dynamic_axes int-key ↔ JSON string-key normalization round-trips through to_dict/from_dict.
• _verify_onnx_export and __post_init__ are aligned: requested dynamic batch logs INFO, unexpected dynamic batch still WARNs.
• Explicit --dynamic-axes supports output tensors too (covered by tests).

Checks: ruff clean on all changed files; full CI green (lint, CodeQL, test commands/models/optim/remaining).

Nice, well-tested feature. Thanks for iterating on the feedback!