LintDiff: Move correlation logic into reconciliation

We're already incorporating data about both "before" and "after" states in the reconciliation code. Use a structure that builds out a set of expected results and then fill that structure with the results as they come in. For each "after" readme/tag match it to a "before" readme/tag. (As it is today) each autorest command for a given readme/tag combination should only be run once.

Start by locating LintDiff's current reconciliation and correlation implementation, then trace how before and after README/tag results are collected and how autorest commands are deduplicated. Refactor toward an expected-results structure that matches each after README/tag with its before counterpart while preserving one command run per combination; done means reconciliation produces the same intended results without duplicate runs.

Written by the indexing model from the issue text.