[tf.data] Add an unbounded thread pool to iterator resources.

The previous implementation of many core `tf.data` transformations
(e.g. `Dataset.prefetch()`) would create one or more threads each time
an iterator over those datasets is created
(e.g. `ds.prefetch(N).repeat(100)` would create and destroy 100
threads). In addition to the overhead of thread creation, this
interacts poorly with some malloc implementations, and can contribute
to memory fragmentation.

The new implementation maintains an unbounded pool of physical threads
in each iterator (or `MultiDeviceIterator`) resource, and returns logical
"threads" to that pool when their work is complete instead of exiting
from them.

PiperOrigin-RevId: 236413014