[Performance] Optimize AdamW GPU kernel and Change the learning rate type to float64 (#78830)

* chore: apply pre-commit format fixes

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

* fix optimizer: add get_lr_dtype()

* fix rtol for opTest

* fix

* fix xpu test and fused_adam_kernel

* fix

* fix win

* fix XPU kernel

* fix

* fix test

* fix sharding.py

* revert bate_pow type

* fix test

* fix test2

* fix test3

---------

Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>