Fix input mutation in branched networks

[mbalesni]

This fixes a bug in networks that have in-place operations (ReLU, ELU, LeakyReLU) right after branching.

Currently, a layer output that is used by several branches can be mutated by an activation function, since they all have inplace=True hardcoded.

Standalone reproduction of the bug:
https://colab.research.google.com/drive/1VBRYou450YNgJdwiDZVRsCfJKUs4FiWh#scrollTo=mKvLPpKz29D

Change: I made the activation functions' inplace property depend on whether the input of the current node has a branching factor higher than one.

Tests: I verified that this passes the tests and does not break conversion of models from download_fixtures.sh.

Example relevant network architecture: Openpilot's supercombo.onnx model:

• the ELU layer (highlighted in red in bottom left corner) mutates its input layer (Flatten)
• another branch that starts at that layer, now gets mutated input, leading to incorrect predictions

Let me know if there is any way I can improve it!