Ensure there are no unncessary whole-file traversals or decodes/encodes

[JukkaL]

Before we freeze the public API, ensure there are no unnecessary performance bottlenecks, such as having to scan through the entire file an extra time, or having to decode/encode the whole file without a good reason. Some of these might impact the public API.

Here are some things that come to mind:

• Do we read the file in both Python and Rust? Can we only do it in Rust? It would be great to only have to read the file once.
• Do we need to scan the whole file an extra time to find # mypy: comments (for flags)? Can we integrate this into the type comment parsing pass?
• Ensure we can avoid semantic analysis pass 1 (tracked separately in Implement all of semantic analysis pass 1 in the extension #11)

[ilevkivskyi]

Note there is one more thing before proper analysis starts: LimitedVariableRenameVisitor() or VariableRenameVisitor(). I think however these should be kept in mypy.