[ARM] Mark i64 and f64 shuffles as Custom for MVE

This way they get lowered through the ARMISD::BUILD_VECTOR, which can
produce more efficient D register moves.

Also helps D115653 not get stuck in a loop.