Implement floating-point manipulation functions for BigFloat

[HertzDevil]

BigFloat already has an implementation of frexp because hash needs it (I think); this PR adds the remaining floating-point manipulation operations.

Methods that take an additional integer accept an Int directly, so this takes care of the allowed conversions in #10907.

BigFloat from GMP doesn't have the notion of signed zeros, so copysign is only an approximation. (As usual, MPFR floats feature signed zeros.)

[straight-shoota]

CI is failing on Windows: https://github.com/crystal-lang/crystal/actions/runs/10698882947/job/29660333970?pr=11007

  1) BigFloat Math .frexp
     Failure/Error: Math.frexp(2.to_big_f ** 1_000_000_000_000).should eq({0.5, 1_000_000_000_001})

       Expected: {0.5, 1000000000001}
            got: {2.43829426113685079336e+327749382, -727379967}

     # spec\std\big\big_float_spec.cr:585

  2) BigFloat Math .logb
     Failure/Error: Math.ilogb(2.to_big_f ** 1_000_000_000_000).should eq(1_000_000_000_000.to_big_f)

       Expected: 1000000000000.0
            got: -727379968

     # spec\std\big\big_float_spec.cr:559

  3) BigFloat Math .ilogb
     Failure/Error: Math.ilogb(2.to_big_f ** 1_000_000_000_000).should eq(1_000_000_000_000)

       Expected: 1000000000000
            got: -727379968

     # spec\std\big\big_float_spec.cr:553

[HertzDevil]

These exponents will overflow even when GMP / MPIR handles the internal exponents correctly, so a fix similar to #14971 is needed. Moreover all BigFloats are always less than 2 ** 0x1f_ffffffc0 when LibGMP::MpExp is 32-bit and LibGMP::MpLimb is 64-bit, anything above that will lead to an overflow in the internal exponent