Optimized _DivideHLBC, fixed incorrect behavior for INT_MIN inputs

[calc84maniac]

Inspired by #595, I took a look at optimizing GraphX's rounded-down signed division routine. Since it wasn't initially designed to round down and the rounding fixup was added later, I thought about ways to make it round down from the start. Eventually, I came up with the idea of transforming the signed division into unsigned, which always rounds down. So, after conditionally negating both the dividend and divisor such that the divisor is positive, the unsigned division can be done as follows:

If x >= 0: Return x / y
If x < 0: Return ((y * 2^24) + x) / y

This setup can be done via an optimized first iteration where the remainder register is initialized to either 0 or y-1 based on the sign bit of the dividend, and that sign bit of the dividend is directly output into the sign bit of the result, allowing the rest of the division to be calculated in 23 iterations.

The one exceptional case is that when the dividend is INT_MIN and gets negated, its sign bit is not actually a sign, so it can't use the same setup. In that case, it simply uses 24 iterations of unsigned division instead.