[Feature Request] Datasets Should Use New `torchvision.io` Image Loader APIs and Return `TVTensor` Images by Default

🚀 The feature

1. Add "torchvision" image loader backend based on new torchvision.io APIs (See: Release Notes v0.20) and enable it by default.
2. VisionDatasets should return TVTensor images by default instead of PIL.Image.

Motivation, pitch

1. TorchVision v0.20 introduces new torchvision.io APIs that enhance its encoding/decoding capabilities.
2. Current VisionDatasets returns PIL.Image by default, but the first step of transforms is usually transforms.ToImage().
3. PIL is slow (See: Pillow-SIMD), especially when compared with new torchvision.io APIs.
4. Current TorchVision image loader backends are based on PIL or accimage, not including new torchvision.io APIs.

Alternatives

1. The return type of datasets can be PIL.Image when using the PIL or the accimage backends, and be TVTensor if using new APIs (may lose consistency).

Additional context

I would like to make a pull request if the community likes this feature.

Read the linked torchvision v0.20 release notes first, then trace the existing PIL and accimage image-loader backends used by VisionDatasets and compare them with the torchvision.io APIs. Done means a default torchvision.io backend is enabled and VisionDatasets return TVTensor images by default, with the PIL and accimage alternatives considered for consistency.