Migrating `skyrl-train` & `skyrl-tx` -> `SkyRL/skyrl` as per tinker-ification

[CharlieFRuan]

As per the tinker-ificaiton effort https://novasky-ai.notion.site/skyrl-tinker, we are re-organizing the repository.

We will merge the skyrl-train and skyrl-tx into the same skyrl folder.

We will keep the original folders (skyrl-train and skyrl-tx) as-is for ~1 week, while copying things to the new skyrl folder, with slightly different organizations.

While developing on top of the repo, please try to develop in the new skyrl folder.

Apologies for the inconvenience!

SkyRL-Train Migration

skyrl_train is now skyrl.train!

Where is the SkyRL-Train code going?

skyrl-train's frontend user facing interfaces/code (trainer/generator, configs,entrypoints,dataset,utils (to do with config validation, trainer utils, etc.), etc...) are now placed in skyrl/train.

Backend logic (distributed,inference_engines/inference_servers,patches,utils (to do with loss, io, etc.), weight_sync, workers, training_batch.py, and env_vars.py) are placed in skyrl/backends/skyrl_train. The code in this folder may undergo further restructuring to remove the skyrl_train naming and make the fsdp and megatron backends stand alone alongside the jax backend that was previously in skyrl-tx.

The skyrl-train/examples, skyrl-train/integrations, and skyrl-train/scripts folders are now at the top level examples/train, examples/train_integrations, and examples/train_scripts.

Migration Plan

The following breaking changes are made as part of this refactoring:

• The vllm extra has been renamed to fsdp
• The main entrypoint is now skyrl.train.entrypoints.main_base instead of skyrl_train.entrypoints.main_base.
• Code structure and imports are now different (as described above)
• The SGLang backend is marked for deprecation and the code path will be deleted shortly. We will consider revisiting it in the future upon refactoring of our inference engine stack: [skyrl][inference] Rollout plan for the new inference backend #1014, and given user interest.