Cache the tf_output object in every tensor lazily on the first call to

_.as_tf_output().
This achieves the extra benefit that while converting the feed_dict of (tensor,
value) to dict(tensor as tf output, value) - earlier we would be adding newly
created objects to the dictionary every time; thereby creating a random order
of feed inputs even when the feed_dict from the client doesn't change.
In turn, this causes a blowup in the number of keys cached by the C++ Executor.
By caching the object, we will end up with a deterministic (yet unsorted) order
while iterating the python dictionary of tf output objects.

As an additional safeguard; we should also consider removing the support of
adding unsorted cache keys to the executors_ in C++ and recommend clients to
use the MakeCallable API (available via make_callable_from_options in python)
if they don't want to have the sorting overhead of cache keys in
GetOrCreateExecutors().

PiperOrigin-RevId: 218784988