Better evals for mock matching algorithm

[jy-tan]

We want to better track the boundaries of the mock matching algorithm and anticipate issues before they arise.

Some thoughts:

1. For psycopg/psycopg2/sqlalchemy spans, we block them from matching on schema because their schema is the same for all kinds of queries, so there's a high chance that schema-based matching will be too inaccurate. Can we more intelligently figure out the structure of the query and only allow variation in specific inputs (e.g., the WHERE clause)? This helps with non-deterministic inputs.
2. Given real spans, can we mutate it and see what will be accepted/rejected?