New generator logic

[tgross35]

We are more or less limited to testing around 2^36 inputs in a reasonable amount of time, which can exhaustively cover unary f32 ops and unary/binary f16 ops, but not much else. I have been thinking with how to better improve what we cover and came up with the following:

• Divide 36 by the number of inputs (36 unary, 18 binary, 12 ternary)
• Divide that number by 4 (9/4/3). Allocate these four bits as follows:
  • One group covers for the sign bit and upper exponent bits
  • One group covers the lower exponent bits
  • One group covers the upper exponent bits
  • One group covers the lower exponent bits
• If the exponent is smaller than 2 times the above number, allocate additional bits to the lower significand bits.
• Exhaustively sweep the assigned bits in these ranges, setting remaining bits to zero.
• Repeat but set the remaining bits to one.
• Repeat but use random values for the remaining bits.

That will cover both negative and positive for zero, max normal, min normal, max subnormal, min subnormal, infinity, qNaN, sNaN, and a number of cases around their transition points, which seems pretty nicely comprehensive.

I think I'd design this as an extensive_edges generator that takes a callback for how to fill the middle portion. This could replace the spaced generator we currently use for extensive tests.

[quaternic]

This reminded me of some code I wrote previously to do something very similar, and I managed to find it in my local clone of libm from before the merge to cb. I can put up a PR for the generator, though it will need additional work to actually use it somewhere.