neon/ld3: load entire vectors sequentially

This should just be slightly easier (less difficult) for compilers
since they may be able to convert this to a gather operation on
some architectures.  I'm not going to hold my breath, though.