Make BigRational.new(BigFloat) exact

[HertzDevil]

BigRational.new(Float::Primitive) is exact, and so is BigFloat.new(BigRational), which GMP provides. BigRational.new(BigFloat) isn't, because it shares the same overload as primitive floats, and assumes only 52 bits of precision are available in the mantissa:

# the default precision is already 64 bits almost everywhere
BigFloat.default_precision = 64

x = 2.to_big_f ** 64
x          # => 1.8446744073709551616e+19
x.to_big_r # => 18446744073709551616

x = 2.to_big_f ** 64 + 1.to_big_f
x          # => 1.8446744073709551617e+19
x.to_big_r # => 18446744073709551616

x = 2.to_big_f ** 64 + 2048.to_big_f
x          # => 1.8446744073709553664e+19
x.to_big_r # => 18446744073709551616

x = 2.to_big_f ** 64 + 4095.to_big_f
x          # => 1.8446744073709555711e+19
x.to_big_r # => 18446744073709551616

x = 2.to_big_f ** 64 + 4096.to_big_f
x          # => 1.8446744073709555712e+19
x.to_big_r # => 18446744073709555712

This PR ensures no bits are lost after using Math.frexp to decompose the argument.