DRAFT: Avoid patching the Keras Model if Keras alrady supports __reduce__

[adriangb]

This patching shouldn't be needed since keras-team/keras#14748 was merged, pending keras-team/keras#15300

[stsievert]

Requiring a higher version of Keras/TF is not a unreasonable approach. Maybe that can be done at some point in the future.

[adriangb]

yep, agreed, that may be a better alternative to this PR. I think next major release may be a good candidate.

[adriangb]

Well I opened keras-team/keras#252 to bump the versions.
Unfortunately, this PR is still not viable because we handle 2 things here which have still not been upstreamed:

• SavedModel in memory on Windows. I think this will be fixed if chore: replace os.path.join with tf.io.gfile.join keras-team/keras#15551 gets merged. This is here in SciKeras.
• SavedModel does not correctly restore optimizer weights (this code in SciKeras, this bug report in Keras).

So for now I am leaning towards bumping the versions in keras-team/keras#252 and then looping back here once/if the above issues are fixed.