add 2x4 tile in mm computation (#4031)

Summary:
Pull Request resolved: https://github.com/pytorch/executorch/pull/4031

The existing optimized mm implementation compute output through 4x4 tile. This isn't efficient when the input tensor's height is a multiple of 3 but not a multiple of 4, e.g. 6. ~~We add a 3x4 tile computation and a parameter `HEIGHT6` to help us choose the computation manner.~~

According to nathanaelsee's experimentation, 2x4 is even more efficient than 3x4, we add 2x4 tile computation and add `TILE_ROW` in yaml files to generate shaders for 2x4 and 4x4 respectively.

Reviewed By: nathanaelsee, liuk22

Differential Revision: D58769774

fbshipit-source-id: 79d8867c87464402b2c6432599b3effc12965122