Add a config to load/save model from/to DCP snapshot

[heslami]

For development purposes, it's often slow to always read safetensor files. Especially for large models such as deepseek v3, we often perform transformations on safetensors (e.g., dequantization and weight mapping) that take a non-trivial amount of time in each run.

So far, folks have been using hacky logic to avoid reading safetensors, relying on specific folder naming structures just so they can save the DCP files in the first run and then use the same written snapshot in subsequent runs.

This PR adds a new config to the policy so that we can optionally pass in the path to DCP. The logic is as follows:

• If the DCP path is given but DCP files aren't present, we download from HF, but after loading the weights we also dump the DCP files.
• If the DCP path is given and the DCP files are present, we short-circuit the run (i.e., we assume that this is a subsequent run from one that loaded from HF and dumped DCP) and directly load the model from DCP.