Fix performance regression

[telser]

A tiny fix, that tested on a closed-source codebase of over 16k modules took formatting inplace from ~10.5 minutes to ~19seconds.

This fixes #1176

[amesgen]

Thanks a lot, a fun/depressing example how the Applicative instance of tuples can be misleading. It probably was introduced while refactoring in #1129 from modifyMVar_ to atomicModifyIORef (where the pure was (correctly) using IO) near the end of the review process. 😬

Essentially, the previous behavior always kept the cache to be the empty Map (as that is mempty for Map which is used for the first tuple component by pure). Semantically, this reverted #915, which we know is very important for performance in large projects (#897).

---

In the future, we might want to introduce automatic benchmarks in every PR that are mandatory to pass before a PR is merged, to avoid such subtle bugs from sneaking in.

[mrkkrp]

Thanks! I will proceed with a new release (0.8.0.2) shortly.

[telser]

Thanks a lot, a fun/depressing example how the Applicative instance of tuples can be misleading. It probably was introduced while refactoring in #1129 from modifyMVar_ to atomicModifyIORef (where the pure was (correctly) using IO) near the end of the review process. 😬

It is a very good example of that!

Thanks! I will proceed with a new release (0.8.0.2) shortly.

Thanks to both of you @mrkkrp and @amesgen!