Handle NaNs when comparing BigInt against Float

[HertzDevil]

Comparing a BigInt against either Float64::NAN or Float32::NAN will raise SIGFPE. This PR fixes that by making #<=> return nil if that is the case. LibGMP.cmp_d already handles infinities, by the way.

Comparisons between BigInts and BigFloats aren't affected, as GMP's BigFloat does not support NaNs (but MPFR's does).

This is a breaking change because previously those return types weren't nilable.