tensorflow: Add some tf.raw_ops members.

[hoel-bagard]

Stubs taken from here.

[JelleZijlstra]

These two must be scalars, so perhaps we can give narrower type (optional)

[hoel-bagard]

I don't have code to run this function, so I can't check, but very often the TensorFlow doc says that the input must be a scalar, but anything that can be converted to a scalar will do. So for example it will say that a float32 is required but "1.2" would work. However that behavior isn't consistent, so we can't know without testing.

For example:

import tensorflow as tf
tf.keras.optimizers.Adam(learning_rate="1.0")   # Raises ValueError
tf.keras.optimizers.Adam(epsilon="1.0")   # Works

So far I think most of the typing for TensorFlow is using overly lax typing for the inputs in order to avoid false positives, so I would leave TensorCompatible.

[github-actions]

According to mypy_primer, this change has no effect on the checked open source code. 🤖🎉