[Feature Request] Shape inference for GroupQueryAttention Op

Nobody has claimed this yet.

### Describe the feature request

For WebNN EP, the graph builder does not accept input and output with dynamic shape. So after FreeDimensionOverride it is expected that all shape / dims are static.
There was already a shape inference function for `GroupQueryAttention` Op in `BaseGroupQueryAttentionTypeAndShapeInference()` of `onnxruntime/core/graph/contrib_ops/bert_defs.cc`. However, the `use_max_past_present_buffer` parameter is set to -1 for each case, as in the following code:
https://github.com/microsoft/onnxruntime/blob/81cd6eacd0121711c2e6e6f8d1fc0cefcd81de99/onnxruntime/core/graph/contrib_ops/bert_defs.cc#L319-L323

So I was wondering if it is possible to pass an argument/flag to give it a chance to perform shape inference, at least when the shared buffer is used by some EPs.

### Describe scenario use case

When some EPs use shared buffer for key / value cache, they pass the flag/argument to set the `use_max_past_present_buffer` to `1`, which will enable the shape inference for GroupQueryAttention Ops.

Start in onnxruntime/core/graph/contrib_ops/bert_defs.cc at BaseGroupQueryAttentionTypeAndShapeInference() and the GroupQueryAttention definitions around the referenced lines. Trace how use_max_past_present_buffer is supplied for shared-buffer cases, then verify that passing 1 enables static input and output shape inference for the affected execution providers.

Written by the indexing model from the issue text.