This avoids potentially infinate loops caused by peeking into
the the Provenance nodes and attempting to copy those references.

These are similar to node_chain_copy() and node_list_copy() which
are there mostly for optimization reasons, but node_copy() and
list_copy() entirely safely separate the nodes from their originals.

This is suitable when giving away a node to a plugin via an API
call where we want to be sure that no modifications modify the source
node.

As a side effect it also ensures that ever node has Provenance, in case
nodes are given back to the data model where they might be expected later.

And use deep copies with both Element.set_public_data() and
Element.get_public_data(), avoiding unintentional mutations of
the underlying data model.

  o None: Calculate cache state if not previously calculated
  o True: Force recalculate cached state, even if already checked
  o False: Only return cached state, never recalculate automatically

And _load_public_data() passes and explicit False value for 'recalculate',
this ensures we never accidentally resolve cached state prematurely if
trying to load the public data as a side effect of calling Element.get_public_data()
outside of the build phase, when all elements in scope should have cached
state resolved and correct anyway.

Plugin assemble() methods may supplement public data returned by
Element.get_public_data() with generated data. Public data is stored in
the artifact cache and automatically loaded as appropriate.

This is too early for dynamic split rules.