feat(perf): show dynamic input dims as dynamic(<actual>) in model info

[xieofxie]

Summary

The pre-benchmark model-info output previously printed dynamic input dimensions as None, which was ambiguous — it didn't convey that the axis was free, nor what size the generated benchmark data actually used.

This changes dynamic dims to render as dynamic(<actual>):

Before

Inputs:      pixel_values         [None, 3, 64, 64]        float32
Outputs:     logits               [None, 1000]

After

Inputs:      pixel_values         [dynamic(1), 3, 64, 64]  float32
Outputs:     logits               [dynamic, 1000]

The value in parentheses is the concrete size the generated input tensor actually used for that axis (e.g. the resolved batch size), so both facts stay visible: the model left the axis free, and the benchmark ran it at size N.

Changes

• Added _format_input_shape(shape, actual) — maps each None dim to dynamic(<actual>), falling back to bare dynamic when no concrete shape is available; static dims are unchanged.
• _print_model_info gained an actual_shapes param, threaded from self._inputs (name → real arr.shape) at the call site.
• Added TestFormatInputShape covering single/multiple dynamic dims, all-static, and the no-actual fallback.

Close #1052

[xieofxie]

tested on

• : dynamic model
• : static model: if not match, will output
  [11:08:54 WARNING winml.modelkit.commands.perf] Requested --batch-size 10 could not be applied: the model's leading input dimension is statically 4. Throughput is scaled by the actual batch (4), not the requested value.