Fix direct memory leak

[hamadodene]

In the current implementation we use PooledByteBufAllocator. PooledByteBufAllocator allows you to have a ByteBuf cache for each thread. So if release is called on a ByteBuf, this is put back in the cache so that it can be reused without having to allocate a new one.

Since it is for Threads, if there are threads that are no longer used and new ones are created this inevitably leads to an OOM. Because unused threads will continue to maintain their ByteBuf cache. We've tried enabling the -Dio.netty.allocator.cacheTrimIntervalMillis option but it doesn't seem to help in any way.

In the way we use the caffeine cache it is convenient to switch to UnpolledByteBuf only for the carapace cache.

The fix consists of:

• Create a UnPooledByteBufAllocator used only for data that needs to be cached
• Each time a chunk is received, the chunk data is copied to a new ByteBuf allocated in UnpooledByteBufAllocator.
• When an entry in the cache is invalidated, the reference Bytebuf is also released so that it can be cleaned by GC.
• Release the original chunk so that the memory portion can be cleaned by the GC.
• Introduce a sys prop cache.allocator.usepooledbytebufallocator=true/false (default false - Unpooled) to use pooled version

To fetch an object from the cache, the logic doesn't change.
It's called retainduplicate() of the chunk.
No need to call release dul duplicate as it will be released already by reactor netty.

Note: reactor netty HttpClient and HttpServer continue use PooledByteBufAllocator. This fix is only to use UnpooledByteBufAllocator for cachable data in caffeine cache.

• I hereby declare this contribution to be licenced under the Apache License Version 2.0, January 2004

[pv3ntur1]

Just a comment

[hamadodene]

@pv3nturi Upgrade reactor netty for potential memory leak in netty 4.1.81.Final polled allocator #424

[pv3ntur1]

LGTM