On-the-fly trace-time DCE using Python ref counting

[dougalm]

We're in the process of replacing jvp+partial-eval with direct linearization which means using DynamicJaxprTrace instead of JaxprTrace to stage the forward tangent computation to a Jaxpr. But JaxprTrace had an unintentional feature that some people have come to rely on: on-the-fly trace-time DCE. Partial eval would build up a DAG of operations instead of a list of eqns, and unreachable parts of the graph would get pruned implicitly. This PR adds the same sort of logic to DynamicJaxprTrace using weak refs so we don't cause a regression.