fix insertion before a node with multiple inputs + support additional broadcasting

[jmduarte]

This PR has a few related fixes

• fix insertion before a node with multiple inputs
• correctly insert broadcast stream before a node with multiple inputs (e.g. Add or Concatenate), e.g. see figure below. The current code would behave incorrectly and replace the wrong input (always the first input) for the following layer... this would result in an additional output incorrectly being inferred
• don't insert broadcast before concatenate (figure below incorrectly does this)

To do

• Add a pytest

@nickshey

[thesps]

Pushed to local branch pr/551 to run Gitlab CI, looks like the graph tests still pass. An addition to test_graph.py that touches this new input_idx argument would be good.

[jmduarte]

@thesps I added the test you requested. It was failing though because the only stream broadcasting we supported was of the form: (1, 1, C) -> (H, W, C), and I was testing (H, W, 1) -> (H, W, C).

So I updated our HLS code to support both of these (easy) cases. I also added tests for both cases and added a RuntimeError if a user tries to use a model with broadcasting that is not one of these two cases.

In principle, it should be straightforward(ish) to also support (1, W, C) -> (H, W, C) and (H, 1, C) -> (H, W, C), but this can be a future PR.

[jmduarte]

@thesps I think this is ready to be merged unless you have any other comments!