Why isn't SDPA supported in cuda within candle?

Hi all, forgive my ignorance but my understanding is that SDPA is cuda compatible but I can't find an implementation within candle_nn for it. I want to run some transformer models (mainly whisper) on T4 GPUs (which don't support flash-attention 2) and want to extract more performance. Optimally, would like something similar to the python transformers library where SDPA is by default used when able.

Start in candle_nn by tracing the existing attention implementations and checking how CUDA-backed paths are selected. Compare the available options for T4 GPUs with the requested SDPA behavior for transformer and Whisper workloads; done means an SDPA path is available when supported and a suitable fallback remains for unsupported hardware.

Written by the indexing model from the issue text.