[hpc][inference] enable cuda graph in engine holder (#66738)
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/66738
added a field `max_batch_size` to TRTModule, which would be later used to determine how big the engine holder would need to pad the input to
Reviewed By: 842974287
Differential Revision: D31286509
fbshipit-source-id: be5c6d4ad9c87ca0842679dc507b187275d4e8dc
