Float/double tests lack NaN and special value coverage

[jonathanpeppers]

The float and double tests only generate finite random values:

rng => (float)(rng.NextDouble() * 2000 - 1000)
rng => rng.NextDouble() * 2000 - 1000

This means the test suite doesn't cover:

• NaN — which is mishandled by the unrolled sort (see Unrolled float/double sorting network mishandles NaN values #10)
• +∞ / -∞
• -0 vs +0

These are known edge cases for floating-point sorting. Tests should verify that NetworkSort.Sort produces the same ordering as Array.Sort for inputs containing these special values, particularly at sizes 27 and 28 where the unrolled path is used.

Found via Copilot code review on #8.