Schedule work away on an unmap, instead of calling it directly. Should result
in less lost samples, and it fixes a lock ordering problem buffer_sem <-> mmap_sem