Support bfloat16, floatE4M3, floatE5M2 types.

[csyonghe]

Problem Description
These types are used extensively in neural network computations. Slang should support them and emit proper SPIRV and CUDA code.

[GinShio]

Hi, if no one has started to implement it, could I take this issue for SPIRV side? Thanks.

[csyonghe]

@GinShio yes, we welcome your contributions! In general, feel free to work on anything that hasn’t been assigned to some one.

[ncelikNV]

I'd like to work on the CUDA side.

[ncelikNV]

@GinShio Did you have a plan for this issue?

[GinShio]

Did you have a plan for this issue?

I guess:

1. add any floating point type conversion in core/slang-math.h.
2. separately implement bfloat16 / floate4m3 / floate5m2 in slang / spirv.

I have an experimental branch GinShio/bfloat16 that simply implements bfloat16. Since it's something I'm personally interested in, it might take a long time to do. I hope to support it before 11/01.

[ncelikNV]

I'll start working on 8-bit types to avoid conflicts with your work.

[GinShio]

Hi, bfloat16_t is implemented in #8757.