Enable fake tensor caching in fbcode by default (#118555)
Summary: Enabled by default in OSS; this switches the default to "on" in fbcode too.
Test Plan: Ran torchbench benchmarks in fbcode
Differential Revision: [D53189048](https://our.internmc.facebook.com/intern/diff/D53189048)
Pull Request resolved: https://github.com/pytorch/pytorch/pull/118555
Approved by: https://github.com/eellison