This shifts the mask register directly, instead of shifting the vector elements and recomputing the mask.
It allows us to operate on a reduced LMUL, instead of the original 4 LMUL=8 instructions, it's now 3 LMUL=1/2, which should be a lot faster: https://godbolt.org/z/5Wc11z7f6

Sadly there are no dedicated mask shift instructions in RVV yet, so we have to emulate it with element slide and element-wise bit-shifts.
While this works, for right shifts it requires vl=vlmax, otherwise the vslide1down may shift in undefined bits.
So I adjusted the code to use the mask shift in the inner loop and handle the tail with the old behavior.

I also moved the implementation::* functions to the respective rvv_*.cpp files to keep the coding style consistent.