Support EvalContext changes

[superdosh]

NodeOutput can now optionally include an updated context. When present, we make sure all parents have the same updated context, and then run the node with that updated context.

• Also had to do some refactoring to avoid circular imports. (NodeOutput -> context.py, rename output.py to verdict.py)
• Key change is around line 180 in dag.py
• Updated tests
• Fixes needed for bijection decoder #122

[github-actions]

MLCommons CLA bot All contributors have signed the MLCommons CLA ✍️ ✅

[bkorycki]

This looks good. I'm wondering if it makes sense to somehow preserve the original context that a node used? For analysis purposes e.g. looking at the outputs from a node should show the context that was actually used at that node's computation time. Otherwise things can be misleading when we look back at the dag run.

[superdosh]

This looks good. I'm wondering if it makes sense to somehow preserve the original context that a node used? For analysis purposes e.g. looking at the outputs from a node should show the context that was actually used at that node's computation time. Otherwise things can be misleading when we look back at the dag run.

This is a good idea. In theory, it should be reconstructible from what's already implemented (since we know the parents of a particular node, we can look up the parents update_ctx, though this is convoluted), so will merge this, but will try to add this shortly more explicitly in a follow up! @bkorycki