[Feature Request] CANNExecutionProvider support for dynamic input shapes.

### Describe the feature request

Currently CANNExecutionProvider always takes a long time to generate om files for different shape inputs.

This is not very friendly for nlp, recognition and other tasks.

Are there any plans to introduce CANN support for [dynamic input shapes](https://www.hiascend.com/document/detail/zh/canncommercial/800/developmentguide/graph/graphdevg/atlasag_25_0054.html)?

### Describe scenario use case

Maybe add input_format, input_shape parameters to Configuration Options.

```python
providers = [
(
"CANNExecutionProvider",
{
"device_id": 0,
"input_format": "NCHW",
"input_shape": "1~6,3,48,-1"
}
)
]
```

Start by tracing the CANNExecutionProvider configuration options and reviewing the linked CANN dynamic input shapes documentation. Determine how provider options such as input_format and input_shape should be represented and validated. Done means dynamic shapes are supported without generating separate OM files for every input shape, with coverage for the new behavior.

Written by the indexing model from the issue text.