[Feature Request] Jagged batches support (NJT) for Transformer / SDPA inference

### Describe the feature request

PyTorch / HF (previously branded as BetterTransformer) now have some support for NJT representation:
- https://github.com/onnx/onnx/issues/6525

This allows to have efficient inference in context of continuous batching.

Does ORT have such NJT-enabled SDPA kernels built-in? (i.e. FlashAttention kernels efficiently supporting block-diagonal masks)

Thanks!

### Describe scenario use case

N/A

No files or tests are named. Start by locating ONNX Runtime's existing SDPA and FlashAttention inference kernels, then compare their mask handling with PyTorch/Hugging Face NJT and block-diagonal-mask support. Done would require a clear determination of built-in support, or a scoped implementation and validation plan for continuous batching.

Written by the indexing model from the issue text.