Consider supporting cached/incremental runs

[RBusarow]

...instead of just doing everything lazily and in-memory.

At the end of parsing a module, serialize all the data collected and write it somewhere in ./build as json or proto. Before parsing, look for the file and inflate to skip the parsing work.

It might still be possible to parse everything lazily. Parse only what's necessary when validating the module's own dependencies, then include a list of "todo" files (the skipped ones) in the serialized artifact. If a downstream module needs to parse more files, it could just parse (some of) the remaining files, then rewrite the serialized summary. This would be more difficult to implement and could easily not be worth it.

If the entire module just gets parsed eagerly, this is simpler in general, but it also aligns with all the extra work necessary in order to validate Anvil graphs.