Patch from Nicolas Pitre

Here's a rewrite of the ARM do_div() implementation.  It is much
faster and smarter than the current code, and it also takes
advantage of ARMv5+ instructions when target processor allows it.

The current code also deserves to be killed ASAP since it overflows
and fails to compute correct values in many cases.  For example:

	u64 n = 2200000001;
	u32 x = 2200000000;
	u32 r = do_div(n, x);

This currently returns n = 41 and r = 46829569 which is obviously bad.

Another failing example is n=15000000000000000000 and x=3000000000.