Fix model parallel device mismatch in `create_bidirectional_sliding_window_mask` (#47560)
* Fix model parallel device mismatch in `create_bidirectional_sliding_window_mask`
#46221 fixed `create_bidirectional_mask` building the mask on
`encoder_hidden_states`' device instead of `inputs_embeds`', but its
sliding-window sibling kept the pre-#46221 code. Mirror the fix and add
a regression test covering both helpers (`meta` stands in for a second
device so the test runs on single-device CI).
* fix
---------
Co-authored-by: vasqu <antonprogamer@gmail.com>