ger implementation, 32/64 split

[jacksonmowry]

Closes: #39

I split out the 32/64 bit implementations but with the way Mojo currently works (at least as far as I can tell) we'll have to also split the tests (unless we want to essentially duplicate all the code within a test function for both precision levels).

I guess the compiler is unable to infer that the comptime dtype parameter is either DType.float32 or DType.float64 meaning passing a buffer of type UnsafePointer[Dtype.float32, ImmutAnyOrigin] is not equivalent to UnsafePointer[dtype, ImmutAnyOrigin] with dtype == DType.float32. Hopefully I'm just missing something and we can go back to having a single test.

This implementation also assumes we want to split the cpu dispatch functions into sger and dger, let me know if we'd like to still just have 1 dispatch function with a dtype template parameter.