[PHI] Major fix for gather/scatter related CUDA kernels (#74922)

* [PHI] gather_scatter kernel largely refactored for correctness

* [PHI] gather scatter kernel rigorously tested

* [PHI] Fixed CUDA 700 error in 4 cases.

5184 forward tests passed, 432 torch comparison failed due to mean int and fp16

* [PHI] Resolve conflicts for scatter/gather kernels

* [PHI] Reformatted with __restrict__

* [PHI] Fix amin smem not allocated bug

Author: Enigmatisms