For the multi device iterator, making sure that the prefetch_buffer_size

doesn't exceed the max_buffer_size.

This was causing deadlocks during shutdown when one iterator (corresponding to say device0) would get destroyed and so that wouldn't be requesting any more data. As a result the MultiDeviceIterator would get blocked on device0 now. device1 might still be alive and requesting data and now its blocked on device0 to fetch (which it won't anymore since its destroyed).

PiperOrigin-RevId: 231456490