Disable fallback for custom kernels (#101131)
Previous failed attempt was here: https://github.com/pytorch/pytorch/pull/97715.
Basically we tried to disable fallback for all ops (aten + custom) but hit many CI failures due to missing fake tensor coverage. Let's just disable it for custom kernels for now.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/101131
Approved by: https://github.com/zou3519
