Decomposition slows down the lazy tensor tracing

compiler
5

This is the version info:

>>> torch_xla.version.__xla_gitrev__
'efa6fcfdac5368330a0770e9019649eba08b5f56'
>>> torch_xla.version.__torch_gitrev__
'f6dfbffb3bb46ada6fe66b5da4f989f9d4d69b3c'