[FEATURE] Potential support for the newer VLA model GR00T N1.7 training and sft

[ShawRong]

Nvidia has published their newer gr00t model with changed architecture.
Here I ask for the support for the gr00t N1.7 which should brought several changes in the training process and configuration.

[nullnonenilNULL]

We've released the loongforge-embodied submodule with support for GR00T N1.7. Please check run_groot_n1_7_ddp_finetune.sh for reference. We are also updating the user documentation shortly.

In addition, our benchmarks show a 50%+ training speedup compared to existing open-source implementations. We will continue to iterate and optimize in the near future. Thanks for your support!