md5: minor optimization in software backend

[newpavlov]

Replaces the logical OR in the G function with addition. It seemingly results in a better ALU utilization and improves performance by several percents. From 699 MB/s to 753 MB/s on my x86 PC and from 910 MB/s to 960 MB/s on Mac M4.

Based on #749

[newpavlov]

Huh... Compiling op_g in isolation results in the same assembly whether we use wrapping_add or not (see here). In other words, the compiler is able to apply such optimization itself. But it seems this change (accidentally?) nudges the compiler towards a better codegen.

[newpavlov]

Interestingly, compiling the full compress function with the wrapping_add change results in more instructions on x86 (672 vs 648), but I guess the resulting code is a bit friendlier to pipeline. On AArch64 the number of instructions is the same (632), but with wrapping_add some orr instructions get replaced with add (i.e. the compiler has failed to apply the optimization when the full function is compiled).

@tarcieri
Could you check whether this change results in a better performance on Mac?

[newpavlov]

On M4 this change improves performance from 910 MB/s to 960 MB/s.

[tarcieri]

Seems about 5-6% faster on my M1 Max