[core] fix DP issue

[younesbelkada]

What does this PR do?

The mini batching PR introduced a small bug that led to DP crash using accelerate as we tried to all reduce tensors that did not had the same shape.
The tensors that did not had the same shape were the logprobs tensors, that derive from the inputs.

The fix is to add a check, that padds correctly the model inputs, and corresponding attention masks, and fills them with the correct value using accelerator.pad_across_processes

cc @lvwerra

run for gpt2 sentiment with DP=2: https://wandb.ai/distill-bloom/trl/runs/p7pzc7yk?workspace=user-younesbelkada
run for t5 sentiment with DP=2: https://wandb.ai/distill-bloom/trl/runs/nb9as16x?workspace=user-younesbelkada

[HuggingFaceDocBuilderDev]

The documentation is not available anymore as the PR was closed or merged.

[lvwerra]

Left two comments, generally looks good.

[younesbelkada]

A CI test (test_ppo_step_with_no_ref_sgd_lr_scheduler) is surprisingly failing but completely unrelated to this PR, I thought it was an issue with PT 2.0 but still cannot reproduce locally. Will merge like this and investigate later