[ROCm] Improvements for vectorized elementwise kernels (pytorch#143269) (#1874)

* Make io_size calculation as minimum of size of input and output size,
rather than the summation of all sizes
* for e.g, for torch.add() on half dtypes (bfloat16/float16),
calc_io_size() returns 6 causing elems_per_thread to be 4
   * But elems_per_thread = 8 works better on half datypes for AMD gpus
* Enable *_load_dwordx4 ISA for 16-bit and 8-bit dtypes on AMD gpus by
using vector size of 8 and 16 respectively

Co-author: @akadutta

Pull Request resolved: pytorch#143269
Approved by: https://github.com/jeffdaily,
https://github.com/pruthvistony

Co-authored-by: Pruthvi Madugundu <[email protected]>
(cherry picked from commit 4686828)

Author: 2 people