Simplify new_rnn_config() by only passing input size, not the full tensor

[hweom]

What does this PR accomplish?

• 🪣 Misc

Changes proposed by this PR:

Replace the src argument to new_rnn_config() with just scalar input size as it's the only information the function needs.

Notes to reviewer:

This is in preparation for reimplementation of RNN layer in new arch.

📜 Checklist

• [x] Test coverage is excellent
• [x] All unit tests pass
• [x] The juice-examples run just fine
• [x] Documentation is thorough, extensive and explicit

[hweom]

Can you clarify? Why would we want to know the specific input data at the point of network construction?

[drahnr]

It's not the input dat, but rather the dimensionality encoded in the container, that a shared tensor really is. And iirc the container is empty by default

[drahnr]

I agree from an API perspective the current state is incoherent