Improved type maximal sharing lookup

[DavyLandman]

Fixes #130.

This simplifies the code and makes it faster in the hot-path.

1. Remove logic that tried to avoid locking while still being incorrect
2. Avoid memory overhead for weakreference lookup for most recent values by keeping around a small map of hot items
3. use Caffeine instead of code added on top of regular maps to simplify logic.

After finishing the migration to java 11, we can test out the performance gains of this change by making a new release of vallang.

Thanks @ben-manes for the discussion and the suggestions in issue #130.

[github-actions]

Unit Test Results

         31 files  ±0           31 suites  ±0   3m 45s ⏱️ +55s
227 506 tests ±0  227 505 ✔️ ±0  1 💤 ±0  0 ❌ ±0 

Results for commit 4c6fd00. ± Comparison against base commit fc8b74d.

♻️ This comment has been updated with latest results.

[jurgenvinju]

ok; the code looks absolutely fine. The non-functional properties I have no idea how to gauge on short notice, so I'm deferring this to @DavyLandman ; we are promising ourselves to evaluate this design in terms of CPU efficiency, thread-safety and memory efficiency in the near future. I'll make an issue.