YJIT: Add codegen for Float arithmetics

[k0kubun]

This speeds up blurhash, matmul, and nbody.

--------  -----------  ----------  ----------  ----------  -------------  ------------
bench     before (ms)  stddev (%)  after (ms)  stddev (%)  after 1st itr  before/after
blurhash  240.8        0.4         233.7       0.4         1.03           1.03
matmul    1026.4       0.1         946.7       0.1         1.07           1.08
nbody     62.6         0.4         55.7        0.3         1.11           1.13
--------  -----------  ----------  ----------  ----------  -------------  ------------

[maximecb]

Your test is perfectly good and valid, but I will point out in case you don't know: numbers having binary fractions are also guaranteed to be representable exactly using floating-point numbers :)

e.g. 1.50, 1.25, 1.125, 2.5 all have exact IEEE FP representations.

[k0kubun]

I wasn't clever enough to figure out 1.125 is 2^0 + 2^(-3) right away, but thanks for a nice refresher 🙂

[maximecb]

This is a nice first step in terms of adding FP support to YJIT 👌

Reading the code, I think it would be good to also add tests that mix fixnum and flonum, e.g. 1.5 * 2