Compilation Cleanup

[beckjake]

Feature

Clean up the compilation process and remove compilation artifacts from the results

Feature description

This is an add-on for #1589.

Currently, compiled nodes store some information on them that varies based on node type (compiled test nodes have all the same fields, so ignore the differences between them for now):

• extra_ctes: This is changed during ref resolution (which happens in the "runtime" context, while compiled_sql is being created). Refs to ephemeral nodes are turned into appends on this array.
• extra_ctes_injected: This is always False until extra_ctes has been processed during compilation.
• compiled_sql: This is None at first, and then is rendered from the runtime context and the raw_sql.
• compiled: This is always True if compiled_sql is a value, and False if it's None. This should be a property, if it even exists.
• injected_sql: This is None at first, and is built from compiled_sql + extra_ctes (+ extra_ctes_injected, but that's just bookkeeping)
• wrapped_sql: This is intially None on all nodes, and on data tests is 'select (count*) from (\n{injected_sql}\n) sbq'. For schema tests and hooks, it is just set to the same as wrapped_sql. Finally, all values of wrapped_sql are passed through rendering. So None becomes 'None'.

I propose we change this so that only the current wrapped_sql (for data tests) and injected_sql (for everything else) is included, under the name compiled_sql, and make everything use that. During compilation we should use a wrapper type of some sort that holds the extra bookkeeping information, instead of attaching that to the actual node spec. We can also remove the cte-related methods from the CompiledNode object, which seems good.

Who will this benefit?

Developers, people looking at the manifest

[cmcarthur]

we should also write accompanying documentation for the various "states" a node can be in -- before this, "parsed", "compiled", "injected", "wrapped" were valid states though the intermediate states ("compiled", "injected") were not actually useful. "parsed", "compiled" seems like a better set of states

[jtcohen6]

With the removal of injected_sql in favor of overwriting compiled_sql (#2762), and standardizing the artifact schemas in v0.19.0, I think we've accomplished the spirit of this issue.

We're still overdue to document the distinction between a node being parsed vs. compiled.