test(e2e): add dynamic-axes e2e tests for export, config, build, perf

[xieofxie]

Summary

Adds end-to-end coverage for the shared --dynamic-axes option, which had no e2e tests. The flag is exposed by four commands — export, config, build, and perf — and this PR exercises each with microsoft/resnet-50 (single pixel_values input, logits output).

--dynamic-axes takes a JSON mapping of tensor name → {axis: symbolic_name} and marks those axes as symbolic (dim_param) in the exported ONNX instead of fixed integers.

What's covered

Each test lives in the existing per-command e2e file, reusing its helpers/fixtures:

• test_export_e2e.py → TestExportDynamicAxes
  • Batch axis becomes symbolic (dim_param="batch") and propagates to logits; channel/spatial dims stay static.
  • Baseline contrast: without the flag all input dims are fixed [1, 3, 224, 224].
  • Multiple symbolic axes (batch / height / width).
  • Invalid spec (empty dim name) fails cleanly with no ONNX written.
• test_config_e2e.py → TestConfigDynamicAxes
  • Mapping is recorded under export.dynamic_axes; absent by default; multi-axis mapping preserved.
• test_build_e2e.py → TestBuildDynamicAxes
  • Symbolic batch dim survives the full export → optimize build pipeline.
• test_perf_e2e.py → TestPerfDynamicAxes
  • --ignore-cache --no-skip-build re-exports, so the benchmarked model reports a dynamic (None) batch axis and the running ONNX carries dim_param="batch".

Note: quantize also has a "dynamic" concept, but that is a quantization mode (dynamic quantization) and not --dynamic-axes, so it is intentionally excluded.

Notes

• Assertions are grounded in observed CLI behavior, not assumptions.
• No hardcoded architecture logic — tests key off the model's real input name pixel_values, consistent with the existing resnet-50 e2e suite.
• All 9 new tests pass locally (uv run pytest -m e2e ...); ruff check is clean.