Maybe add half (16-bit float) support

[ulfjack]

TBD

[ulfjack]

This is actually fairly easy - we can just reuse the 32-bit float computation. Except for the decoding, the code is independent of the original bit size. I've experimentally implemented 32-bit via the 64-bit conversion process and it passed all the tests.

[ulfjack]

Alternatively, we can use the new generic API, which should cover everything up to and including 128-bit float.

[ulfjack]

(Although not with quite the same level of performance.)

[quinnj]

Just a note here that the Julia implementation supports Float16; supporting it was pretty straightforward, following the Float32/Float64 patterns, but I did end up porting the split-table generation code to Julia and we just generate the tables during module precompilation, in addition to generating the Float16 tables.