compare ints against doubles without rounding

[pranit0808]

1. compareDoubleInt and compareDoubleUint converted the integer operand to a double before comparing, so any int or uint above 2^53 was rounded onto a neighbouring value first.
2. Equal is defined as Compare() == 0 for all three numeric types, so int(9007199254740993) == 9007199254740992.0 evaluated to true and MaxInt64 < 9223372036854775808.0 evaluated to false.
3. That also disagreed with map and list indexing, which already resolve cross-type numeric keys through the lossless helpers in overflow.go: with those same two values x in [d] returned true while {d: 1}[x] reported no such key.

Split the double instead of converting the integer: range-check it, compare the integral part directly against the integer operand, and let the remaining fraction break the tie. NaN never reaches either helper, every caller screens it first.