[NVIDIA] Refine API difference of all_gather_object

[anderson101866]

Unlike PyTorch, using paddle don't need to pre-allocate object_list to equal size of group size.

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}