Explore GPU/NPU support

[mkopcins]

This task focuses on adding support for GPU / NPU in our exported models.

As a result of this task there will be models exported utilizing GPU/NPU supports and issues that blocked us from exporting:

Whisper:

• XNNPACK
• CoreML
• Vulkan - Add support for torch.complex64 on vulkan backend pytorch/executorch#17791

YOLO all variants:

• XNNPACK
• CoreML
• Vulkan

Selfie segmentation

• XNNPACK
• CoreML - This is probably not needed
• Vulkan - This is probably not needed

RF-DETR

• XNNPACK
• CoreML Norbert exported, need to add it
• Vulkan

LFM2.5 Instruct:

• XNNPACK
• CoreML
• Vulkan

Style transfer models:

• XNNPACK
• CoreML
• Vulkan

sddlite-320:

• XNNPACK
• CoreML
• Vulkan

Image embedding (CLIP):

• XNNPACK
• CoreML
• Vulkan

efficientnet-v2-s:

• XNNPACK
• CoreML
• Vulkan

Interesting refs:

• https://pytorch.org/blog/building-voice-agents-with-executorch-a-cross-platform-foundation-for-on-device-audio/