Do not use warm_pool() if TorchTnT is used (#121047)
Summary: This diff is needed to avoid QPS drop when parallel compilation is used with TorchTNT.
Test Plan:
On TNT
* https://www.internalfb.com/mast/job/torchx-ldm_train-hxjhl0k1wjz93
On PyPer
* f537224855
Differential Revision: D54430900
Pull Request resolved: https://github.com/pytorch/pytorch/pull/121047
Approved by: https://github.com/yanboliang