This is a fix of reverted #646

• Extract shl/shr by a limb-sized shift into separate methods
• Add a more detailed docstring and some debug assertions in div3by2()
• Replaced saturating_sub with wrapping_sub in several places. While logically it is the same thing (the wrapping/saturation only happens for values that are later selected out), I think there are readability advantages. First, elsewhere in the code we use wrapping ops for selected out values (meaning "perform the subtraction without any checks, since we already have a constant-time condition for that"), so saturating_sub indicates that the algorithm actually uses the saturation mechanic. Second, in case of a bug, it will be easier to spot the consequences of a "0xffff..." value than a 0.

Update:

• Use div2by1 instead of div3by2 with zero lower limbs. That's why I was wondering if div3by2 should be moved out of div_limb.rs — when it's there it's perceived as a general-purpose method analogous to div2by1, but it's not - it may return the quotient or the quotient plus 1. It is only supposed to be used in a specific place in the division algorithm. It returns a WideWord as a remainder, which was truncated to Word by the caller without any checks or explanations why it's supposed to be correct.