[iOS GPU][Design] Use function_constants to simply shader kernels (#54518)
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/54518
When I was reading the Metal Shader Language Specification, I noticed that using `function_constants` in C++ attributes could let us do compile time kernel selection, which can drastically reduce the complexity of writing GPU kernels for different input texture types. We should apply this trick to all our existing shader functions.
ghstack-source-id: 124497703
Test Plan:
- Metal op tests
```
2021-03-20 23:35:20.496922-0700 PyTorchPlayground[48215:8455407] [bool test_view()],[1 10 2 2 ],[SUCCEED]
2021-03-20 23:35:20.522714-0700 PyTorchPlayground[48215:8455407] [bool test_view2()],[1 10 2 2 ],[SUCCEED]
2021-03-20 23:35:20.553591-0700 PyTorchPlayground[48215:8455407] [bool test_view3()],[5 8 ],[SUCCEED]
2021-03-20 23:35:20.571194-0700 PyTorchPlayground[48215:8455407] [bool test_view4()],[5 8 ],[SUCCEED]
```
- Sandcastle CI
- CircleCI
Reviewed By: SS-JIA
Differential Revision: D27218965
fbshipit-source-id: 763c54d551de3a88e4ff0007894200d72f00958c
