[tf.data] Compute short circuit indices at DatasetOpKernel construction time.

This change optimizes the creation of nested datasets (e.g. inside `Dataset.interleave()`) that use short-circuitable functional transformations (e.g. `Dataset.map()`), by instantiating the function once at kernel construction time to determine the short-circuit indices, and caching the result in the kernel.

PiperOrigin-RevId: 237073994