Performance with only F16C extension

[eqrion]

Motivated by #7, what is the expected slow-down for the F16C-only lowerings?

It seems like the majority of our users would be using this path.

[dtig]

The expected slow down for using pure emulation is quite significant, but hard to quantify because it depends on the exact workloads being measured and the extent of the numerics used there. The recommendation on older hardware would be to either use accumulator kernels that use conversions at the boundary, and processing in f32, or default to detecting fP16 support and use a different code path.