[tf.contrib.data] Avoid creating threads in the ParallelMap dataset.

This existing `ParallelMap(f, num_threads, output_buffer_size)`
dataset has an iterator that uses `num_threads` threads to invoke
`f`in parallel, and prefetch elements into a bounded buffer of max
length `output_buffer_size`. This change modifies the implementation
in two ways:

* It replaces the `num_threads` worker threads with an equivalent number
  of concurrent async `Run()` calls on the underlying
  `FunctionLibraryRuntime`.

* It factors out the prefetching behavior from ParallelMap, into a separate
  Prefetch dataset, as suggested by jsimsa@.

To reflect the change in implementation, the `num_threads` input to
`ParallelMapDatasetOp` is renamed to the more abstract
`num_parallel_calls`.

The net result is a reduction in the number of additional threads per
parallel map from `num_threads` to 1 (if `output_buffer_size` is
specified) or 0 (if `output_buffer_size` is not specified). Since
these background threads offloaded work to the existing inter-op
threadpool, there is no effective reduction in parallelism, and this
approach should enable us to scale to larger degrees of parallelism
because the system will have fewer threads to manage.

PiperOrigin-RevId: 166154782