Refine type restriction of Math.frexp(BigFloat)

[straight-shoota]

The return type of this overload depended on LibC::SizeT. This patch unifies it to {BigFloat, Int32} on any architecture.

This is technically a breaking change. Restricting the union in the return type restriction further down to one member of the union is technically fine. But the union did not happen in practice, because the actual type was always Int32 or Int64 depending on the architecture.

#10645 (comment)

[HertzDevil]

On 64-bit systems BigFloats can get large enough to have exponents that don't fit into an Int32, so I'd prefer to_i64 over to_i32:

Math.frexp(2.to_big_f ** 1_000_000_000_000_i64) # => {0.5, 1000000000001}