geohash is a stable encoding algorithm. It should be a one time transformation of data point to a geohash value. Squeezing out absolute performance of the geohash algorithm is not practical use of time and resource.
But you see, the author described this absurd optimisation as fun. Fun is a very practical use of time and resource!
You're right: No fun is EVER allowed. Is that a smile on your face? Doesn't look very productive to me, now get back to work.
What if I need to do this to a large dataset that hasn't been geohashed before?
You run it once, and on each new data element you introduce into the collection.
Or if you prefer, geohash it JIT style, and once it's persisted, you don't run it again thereafter.
If I have a billion of them to hash, it's still worth it to think about how to do the hashing more efficiently. Just because it's not something I need to do to a large dataset constantly doesn't mean that it's not worth thinking about how to do it efficiently because I only have to do it once.