Support lowering quantized checkpoint from HF Hub

[guangy10]

Users can lower a HF model with quantization by:

1. Load the fp32 checkpoint and apply the quantization, export and lowering recipe to generate the PTE, which is supported in Introduce 8da4w quant for decoder-only text models #62
2. Load a pre-quantized checkpoint and directly lower it to ExecuTorch via the exact same from_pretrained API. Behind the scene, it will bypass the quantization as the checkpoint has been quantized already. This path shows the composability of different recipes.

This PR is handling path 2 above.

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[guangy10]

E           ValueError: Failed to find class AOPerModuleConfig in any of the allowed modules: torchao.quantization, torchao.sparsity.sparse_api, torchao.prototype.quantization

Will need to upgrade the pinned torchao.

[guangy10]

Rebased after #68 is merged

[guangy10]

cc @metascroy @jerryzh168 for review.