[AArch64] Improve lowering of insert_vector_elt with 0.0 consts.

When moving +0.0 into a float vector, we can use to vi*gpr variants of
INS.

Reviewed By: efriedma

Differential Revision: https://reviews.llvm.org/D90176