[BREAKING] Fix for OOM error when reading from S3

[malhotrashivam]

Context:
In the current code, each S3Context object (which is created per thread and stores any data to be used by that thread for its operations) has a cache storing hard references to the last few fragments that this thread has fetched from S3. By default, the cache size is 256. This was added to prevent recently fetched fragments from getting freed by the garbage collector. The rest of the fragments are referenced via a soft-reference based cache, shared across all threads.

Problem:
If a user configures really large fragment sizes, like 20 MB, then the local cache size grows very large and since these are all hard references, they cannot be freed by the garbage collector and can potentially lead to an Out of Memory error.

PR description:
As part of this PR, we have removed the thread local cache of hard references, and all the caching will be done via a soft reference based cache shared across all threads.

Breaking:
We have removed the cache related s3 specific instructions from the API, namely S3Instructions#maxCacheSize() from java and s3.S3Instructions.max_cache_size from python wrapper.

[chipkent]

Python LGTM

[rcaudy]

We discussed some minor changes, but I'm ready to approve this.