Directly supporting Megatron Core checkpoints

Please 👍 this issue if this if you'd like to see us support this.

The checkpoint format between megatron and the format we use is very very close so we should be able to support a thin translation for the metadata. The obvious limitation with this is we wouldn't be able to support vllm in general due to there not being an HF and/or vllm implementation of the model; so the user can only use the megatron inference backend.

Related #619