[LTC] Use DataCache in GetIrValueForScalarFromCodegen
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
