[LTC] Use DataCache in GetIrValueForScalarFromCodegen (#92066)
Summary:
XLA expects GetIrValueForScalarFromCodegen to use DataCache such that not every scalar will request a data transfer to the backend device.
This needs pytorch/xla#4447 to verify.
Test Plan:
PJRT_DEVICE=CPU python xla/test/test_operations.py -v -k test_cached_addcdiv
Fixes pytorch/xla#4213.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/92066
Approved by: https://github.com/JackCaoG
