[PyTorch] Fix minor bug in computing num_gqa_groups_per_partition

[knowlsie]

Currently the number of GQA groups per partition for DotProductAttention is computed using a default value, rather than the actual value computed earlier in the initializer, causing errors when tensor parallelism is enabled. This just fixes that.

[ksivaman]

Thanks for spotting this! @knowlsie

[ksivaman]

/te-ci pytorch

[knowlsie]

Thanks for the quick review @ksivaman — I can’t merge without write access so feel free to merge whenever.

[ksivaman]

Fixed in #1044