[RISCV][InsertVSETVLI] Avoid VL toggles for extractelement patterns

Noticed this while looking at some SLP output. If we have an extractelement, we're probably using a slidedown into an destination with no contents. Given this, we can allow the slideup to use a larger VL and clobber tail elements of the destination vector. Doing this allows us to avoid vsetvli toggles in many fixed length vector examples.

Differential Revision: https://reviews.llvm.org/D148834