[clang] [Matrix][Clang][HLSL] Move MaxMatrixDimension to a LangOpt (PR #163307)

[Chris B via cfe-commits]

llvm-beanz wrote:

The behavior change is that in HLSL-mode the `matrix_type` attribute has reduced size limits.

> We are confirming that the constraint is applied on matrix<int, 5, 5> mat3;
> 
> https://github.com/llvm/llvm-project/blob/b358af10a7b783e771a1e5ad6ded3e3535648b06/clang/test/SemaHLSL/BuiltIns/matrix-basic_types-errors.hlsl#L10-L12

This is enforced by the `matrix` template, you're changing the implementation to _also_ enforce this on the `matrix_type` attribute.

> Since the behavior differences are tested. I’ve just added an explicit compiler invocation test to confirm this number changes When we change the language mode to HLSL. I think that’s all you wanted?

With this PR, clang should generate an error on this code:
```c++
using float8x8 = __attribute__((matrix_type(8,8))) float;
```

Which would not have been generated without this PR (See [Compiler Explorer](https://godbolt.org/z/Gh53nEGEs))

https://github.com/llvm/llvm-project/pull/163307