Try optimizing f.mvs by only allocating when the length changes

[kkysen]

I think most frames should be the same size, so hopefully this helps enough without having to have a whole pool allocator, which gets tricky with lifetimes and self-referential structs.

@ivanloz, could you look at this version, too? Since this one doesn't use any unstable features.

[ivanloz]

I mentioned it in #1358 (comment), but I didn't see any difference with this patch applied unfortunately.

[fbossen]

Isn't the memory released anyway at the end of each frame in rav1d_decode_frame_exit (see let _ = mem::take(&mut f.mvs)) and therefore reallocated in rav1d_submit_frame?

[kkysen]

Isn't the memory released anyway at the end of each frame in rav1d_decode_frame_exit (see let _ = mem::take(&mut f.mvs)) and therefore reallocated in rav1d_submit_frame?

Hmm, I didn't think of that.