Use __ldg for CUDA kernels in fuser (#18540)

Use __ldg for CUDA kernels in fuser (#18540)

Summary:
While benchmarking a kernel with broadcasted inputs, I noticed
that is was much slower than a hand-coded kernel for the smae task.

The kernel in question computed a * b + c for a of shape
32 x 32 x 10240 and b and c of shape 1 x 32 x 1.

This patch accellerates said kernel from 450us to 250us on my GTX1080Ti.

I didn't change half because there doesn't seem to be __ldg for
half.

An alternative could be to sprinkle const and restrict.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/18540

Differential Revision: D14657840

Pulled By: soumith

fbshipit-source-id: 408847346ec12d1d1d9b119ac50bbc70f0d9ed33