vmx: Reimplement create_mask_32_128 and use it in vmx_fill

Based on suggestion from @siamashka.

This lets the compiler pick the vector instruction to use which is
usually the best idea.

Use create_mask_32_128() instead of create_mask_1x32_128() in
vmx_fill(), avoiding loading memory beyond the filler argument on the
stack.

Remove the now-unused create_mask_1x32_128(). This gets rid of some
(correct) warnings from the compiler about indexing beyond the variable
in question.