[GPT-OSS] Add HF state dict adapter to support loading from HF checkpoints

[shuhuayu]

As titled, this PR adds HF state dict adapter to support loading from GPT-OSS HF checkpoint. GPT-OSS checkpoint is quantized in MXPF4 format. The de-quantization steps are offloaded to the QuantizedHuggingFaceStorageReader in dcp, so this feature depends on this PR to update QuantizedHuggingFaceStorageReader (pytorch/pytorch#167672).

1. Test 1. We use dcp.load(hf_state_dict, storage_reader=QuantizedHuggingFaceStorageReader(path=input_dir)) to load from GPT-OSS HF checkpoint, and map the hf_state_dict back to TorchTitan state dict. We build one test input, and compare two outputs: 1. Using transformer library to load GPT-OSS HF checkpoint and run inference on the test input; 2. We use the converted TorchTitan model to run inference on the test input. We compare the outputs by comparing the KL divergence of two output probability distributions. The result shows two models are very similar.

2. Test 2. We load the model directly from quantized GPT-OSS HF checkpoint, and do a test training.
   

[tianyu-l]

SGTM!