Improve extraction of double vectors from quad vectors

There are NEON patterns to extrract one half of a quad vector. These are
presented at the rtl level as instructions with arbitrary quad sources and
arbitrary double destinations. If the register allocators happen to assign
a good combination, the patterns expand to nothing, otherwise they
expand to a vector move. However, there is nothing to tell the register
allocator what is required here.

In principle, we should be able to use a plain subreg move instead.
That seems to trigger some latent middle-end bugs though.

Fixing this removes several vmovs from some libav loops.

Status:

Patch merged to 4.6.