Create generic floating-point wrapping type for <NumExponentBits, NumMantissaBits>

[jrhemstad]

All the different extended floating point types like __half, bfloat, fp8, fp4, fp6, etc. are basically just a collection of N bits where M bits are dedicated to mantissa, E bits to exponent.

We can lean into this idea by creating an overarching type that is just:

floating_point<NumMantissaBits, NumExponentBits>

this would dispatch to corresponding built-in or accelerated types for the appropriate number of bits.

Tasks

Beta Give feedback

1. Provide generic implementation of floating_point<M, E> limited to arithmetic operations #1666
   +0
   
2. Create new half_t type that aliases floating_point<5, 10>
3. Make floating_point<5,10> dispatch to accelerated instructions where possible
4. Create bfloat_t type that aliases floating_point<8, 7>
5. Make floating_point<8,7> dispatch to accelerated instructions where possible