Roll back support for caching encoding detection
This feature can lead to correctness issues with binary detection (see #340013 (closed)) since blob data is not guaranteed to be loaded entirely when detection is run.
Benchmarks in #34001 (closed) show on staging that the cost of Redis lookups can be 80x slower than just running Charlock Holmes of up to 8000 characters per file.
Relates to:
Edited by Stan Hu