DNM: Handle pipeline breakers through avoiding reuse

[phofl]

This is a naive implementation that recomputes pipeline breakers (#854) to avoid OOM errors. This PR has a list of deficiencies, namely:

• We always go back to IO Nodes, anything that shuffles before we run into a reduction should be fine
• We should track references on the Expression, this would mean that the pipeline breaker modifications would be a simple "_simplify_down`` step. We are messing with nodes deep down in the tree at the moment which means that their dependents aren't updated properly. That's the main reason that it currently is a separate optimization step
• Not all reductions are supported
• I think we have to lower merges into different expressions to make this work properly

Just putting this up here for others to potentially mess around with. This is most helpful for query 21 in tpch benchmarks. The query is super slow without dask/dask#10922 though