babl: _mm_load_ps() requires a 16-bytes aligned address.

Without this, babl_process() requires the source pointer to be 16-bytes aligned (hence we should document it). But if we don't want to add such additional requirement, then we must add the additional test and optionally (in case it's not 16-bytes aligned), move to an address which we know to be aligned.

Of course, this case implies an additional memcpy() so it will be a bit less efficient, but at least babl_process() doesn't crash and we keep its usage generic.

See also: gegl!142 (merged)