[REQUEST] make ZeRO inference config more intuitive wrt `train_*` configuration

With the great improvements done over the last few months ZeRO can be now be easily used for inference, as in this example:
https://github.com/huggingface/transformers/issues/15399#issuecomment-1025240005
But to the user who may no longer care for training it's no longer intuitive to have these train-specific config variables:

    "train_batch_size": train_batch_size,
    "train_micro_batch_size_per_gpu": 1,

it's time to come up with a train/inference-non-specific config names and deprecate the train-specific ones.

there might be others, but these 2 just stood out.

Thank you.

Start by locating the ZeRO configuration handling and references to train_batch_size and train_micro_batch_size_per_gpu. Review how configuration deprecations are tested, then define inference-neutral names, deprecation behavior for the existing names, and coverage for both training and inference configurations.