[tf.data] Avoid copying `std::shared_ptr<model::Model>` on hot paths.

This improves the performance of the `Dataset.range()` benchmark with
and without modeling enabled:

* Average time per element (without modeling): 110.381937 nanoseconds
  (was 124.318142 nanoseconds)
* Average time per element (with modeling): 498.648620 nanoseconds
  (was 584.610391 nanoseconds)

PiperOrigin-RevId: 220680272