copy src into dst (the shorter length) with non-temporal stores where the target has them
(x86-64: the 16-byte aligned part of dst): dst is neither read first nor pulled into the
caches, which keep what the thread works on. Ends with the fence the stores need (sfence)
before anything touches dst again; copy_stream_rows fences a block of rows once.
copy_stream for rows rows of width bytes: row y of src starts at y * src_stride, of dst
at y * dst_stride; one fence at the end. Panics when a row lies outside its slice.
the streaming stores of this thread become visible to others (sfence); copy_stream and
copy_stream_rows run it themselves. Nothing to do on other targets: without such stores the
release of whatever hands dst on is enough