Use Caffeine for request aggregation

[SailReal]

As we found out and demonstrated using GuavaCacheTest.java that the Guava-Cache isn't completely thread-safe, we're switching to Caffeine for metadata request deduplication:

The new MetadataRequestDeduplicationDecorator decorates an existing CloudProvider by de-duplicating identical itemMetadata and list-requests so that the delegate is called only once until the future is completed.
Furthermore, quota-requests are cached for a duration of default 10 seconds (can be set using org.cryptomator.cloudaccess.metadatacachingprovider.timeoutSecond).

In projects that use this library, we will write a factory that selects the best metadata decorator, currently MetadataCachingProviderDecorator or MetadataRequestDeduplicationDecorator.

[overheadhunter]

Great work so far. Just some small suggestions

[ben-manes]

As we found out and demonstrated using GuavaCacheTest.java that the Guava-Cache isn't completely thread-safe, we're switching to Caffeine for metadata request deduplication:

I think you mean that invalidate(key) does not discard an in-flight load (google/guava#1881)? That is thread-safe but not linearizable, which can cause surprises if one depends on that to avoid stale loads. Caffeine does support that (and verified by Lincheck) for keyed operations.

Unfortunately we cannot natively do that for invalidateAll() / Map.clear() due to in-flight loads being suppressed by ConcurrentHashMap's iterators (which guarantee visibility of inserted mappings). As we need to maintain our own metadata, linearizable bulk operations would require forking the hash table for internal access (see CHM's clear() implementation). That caveat is simple to workaround for AsyncCache since the load is performed by the future, allowing the mapping to be established immediately and one can easily rework the logic to maintain linearizable clear() semantics. If that is of interest / need for your usage, we can discuss that further. In short, you would use either JCIP-style memoization instead of computeIfAbsent style, or have the load wait on a flag to force the computation to starts after mapping was established, or include a generational id as part of the key.