Invalidate outputs based on better than timestamp

[natebosch]

Today a changed file can cascade to any output which may end up rebuilding way more than is optimal. If we track some type of content hash we can get away with pruning the build graph more eagerly.

[jakemac53]

This will also help for correctness, so I am going to go ahead and tackle it.

See #536 for an example.

[jakemac53]

Actually I am only going to tackle invalidating inputs, not also invaliding outputs based on whether files actually changed (at least right now).