Previously it was more an all-or-nothing where either there
were values provided from the index or there wasn't.
With the native composite index there may be value tuples
which contain some NO_VALUE and some actual values.
Therefore the implementation of acceptNode has been changed
to accommodate for it.

Where it could generate keys that could direct tree traversals
in the wrong direction.

The reason was that the minimal splitter logic would first find
which key slot (out of potentially many, for composite) would
be the first to differ - copy those before it, generate
minimal splitter for the first differing slot and then initialize
the rest to "dummy" values. Problem was that the dummy value was
actually a valid arbitrary value (zero integer to be exact)
and would therefore affect comparisons in ways it wasn't
intended to do.

This commit changes this so that the first differing slot AND
all the other slots AFTER it generate minimal splitters.

After each accessor/populator test

getCause to be a root cause.

In cases where BackupProtocolService was created without externally provided
pageCache another separate independent instance was created and it was need closed.
In addition to leaked page cache memory, the scheduler that independent page cache
was using was also never closed and that was the source of leaking threads.

To be able to avoid memory and thread leaks new
BackupProtocolServiceFactory was introduced to allow constructions of
various variants of BackupProtocolService.
Also service class itself updated to be AutoCloseable.
As result now BackupProtocolService clients have the ability to release all
resources by closing service after required operations.

To be more concise and more easily understood.

Allow to lookup dynamic address of Prometheus server. Update test to use
newly introduced capability.

The Netty dependent Channels that is used in causal clustering now
implements WritableChannel. Before they implemented FlushableChannel
for no apparent reason. They have also been renamed.

Marshalling for RaftV1 broke throughout development. By using the now
introduced ByteBufBacked interface, the tx marhalling for RaftV1 can be
optimized, writing directly to ioBuffer.

* Separate marshalling and codec to more clearly expose the difference.
 RaftV1 uses marshall for everything but in RaftV2 codec is used over
 the network.

* Use a visitor pattern when marshalling replicated content.

* Move tx serializeation logic back to the
 ReplicatedTransactionSerialization class and the transaction writer to
 ReplicateTransactionFactory.

This reverts commit 20ab5e7be0c02f1b15f3bdc16499bddba6c4d8af.

Resolve review comments regarding formating