Add more HF models with bug fixes.

[garrett4wade]

1. Add a manual PPO script example.
2. Add deterministic seeding in pytorch dataloader to increase reproducibility.
3. Fix a bug in partitioning micro-batches of paired preference data in the pipe runner. (Related to RW and DPO training).
4. Support several new models: QWen2, Gemma, and GPT-2. ReaLModel and its configuration are changed correspondingly to support tied embeddings .
5. Support loading safetensor HF checkpoints.
6. Fix the mapping of contiguous parameters to make the Megatron backend happy.
7. Fix the finalize_grad function of the Megatron backend to avoid blocking when training with a partition of all GPUs.
8. Fix gradient all-reduce for LayerNorm and tied embeddings.
9. Add inference tests for new models.
10. Add a parameter reallocation test.

Known issues:

• The correctness of parameter reallocation is ensured with more assertions at the cost of efficiency. The issue should be resolved with further profiling and implementation.
• Should check and update the docs.

[nuzant]

LGTM, only some minor changes required.