Sumdigits exmaple with GRPO

[Ritesh1905]

The reward-mean converges in less than 15 training steps and takes ~5 mins

wandb run: https://meta.wandb.io/torchforge/sumdigits-training/runs/uxzowpkp?nw=nwuserrithesh

• Made some corrections to GRPO loss function.
• Updates the sumdigits example to use GRPO loss
• Introduces a small curriculum setup.

TODO:

• Lot of code redundancy, which will be fixed once the abstractions are landed.