[CORE] Refactor XferCRC to make it branchless and to remove winsock dependency

[Mauller]

This PR refactors the XferCRC to make the code branchless and to remove the winsock dependency within it.

There may be minimal performance improvement from this as other factors will need changing in the data being passed to the CRC to further improve the performance.

The winsock dependency has been replaced with the endian compat library instead.

[Caball009]

Could try a switch statement to unroll the loop, though it'd have to be measured to see whether it's an improvement:

switch(leftover)
{
case 3:
	val += (c[2] << (2 * 8));
case 2:
	val += (c[1] << (1 * 8));
case 1:
	val += (c[0] << (0 * 8));
default:
	break;
}

[Mauller]

I will check this out, but if it works that at least gets rid of the second loop.

[Caball009]

Perhaps we could go for even fewer branches by using the data size as a template parameter. Most of the call sites use a size that's known at compile-time.

[xezon]

Replacing the winsock call is good. Removing the hibit branch also makes sense, because the branch predictor likely does not work well there (50% chance). I would like to see a performance comparison when this works without any remaining logical mismatch.

[Mauller]

Fixed the logic on validity handling of the leftover bytes.