Whisper - add CoreML & Vulkan backends support

[IgorSwat]

Problem Description

Right now, all the Whisper models are utilized only within the XNNPACK backend. We could extend the support by exporting them to Vulkan & CoreML backends.

What should be done

Export Whisper family models to CoreML & Vulkan if possible.

Benefits to React Native ExecuTorch

Significant speed-up in speech to text module. A possibility of using bigger model (for example: whisper-base) with similar performance but improved accuracy compared to currently standard whisper-tiny.

[msluszniak]

https://github.com/software-mansion/react-native-executorch/tree/%40ms/experiment-with-npu-and-gpu for a reference for coreml

[IgorSwat]

Current status:

• CoreML backend: models exported succesfully, resulting in a huge performance boost
• Vulkan backend: models export properly, but fail during runtime inference with libc++abi: terminating due to uncaught exception of type std::out_of_range: vector. The problem occurs in preprocessing phase of the inference and a possible cause might be Vulkan's limitation on a vector size (initial audio vector contains hundreds of thousands of samples).