Microoptimize integer-length for signed-byte 32 and unsigned-byte 32 args

When implementing the fix for #283 (closed), we take advantage of the fact that BSR does not modify the destination register when the source is 0. We can optimize the signed-byte-32-int-len vop using this fact too. It gets rid of an extra jump and an extra instruction.

We can also optimize the unsigned-byte 32 version by assuming that the BSR operation copies the source value to the destination register when the source is 0. Gcc, LLVM, and MSVC generate code assuming this happens.