Adjust MD5Transform in line with existing proposals for SHA1Transform.

Make i386 version nostackframe and save some pushes. This might even fix it for Darwin (unsure).

Also, remove the so-called “original version”. Its only point is being a slightly more readable reference, but the C/++ sources that are designed around the fact R1..4 are macros are better suited to convey the related knowledge; and there can be no reason to have its derivative enabled specifically on x86-64 rather than everywhere. (Does it presently ignore endianness btw?.. Okay, so maybe this was the reason, I added a conversion.)