Change arg dtype from float to double in LPPool and nn/utils/clip_grad.h (#29584)
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/29584
In Python, `float` dtype is always 64-bit (https://stackoverflow.com/a/8216110), and the C++ equivalent APIs should take `double` dtype to match the bit length.
Test Plan: Imported from OSS
Differential Revision: D18436616
Pulled By: yf225
fbshipit-source-id: ece510bba6f089ccada03af216f4805bbd03f5f2