New model for writing records to file

[mkeskells]

The current implementation buffers records, and writes them based on a flush

A more appropriate approach is IMHO to stream the records to file
In this model the RecordGrouper would not collect the records in the buffer, but indicate the logical stream of where the records should be written, and the caller ( the task) manages how the records are handled - added to a "logical stream"

This opens up several possibilities and improvements

• files can be written when the logical stream has sufficient buffered recordsl, rather than waiting for a flush
• files could be written in parts when there is sufficient data to write, but the file isn't full
• memory pressure is reduced when records are written (maybe only after the file is closed)
• the IO load is more distributed, (not maybe several hundred files being written when a flush occurs)

other possibilities

• files can be written asynchronously, (ideally in virtual threads when we are building on a JVM that supports that) which makes the system more responsive (reduction in latency)
• we can make a better back pressure model, that is more graceful
• back pressure can be graceful, rather than a "stop the world action"

We can use requestFlush() when files have been written to update the metrics, rather than relying on timeouts and flushing everything

Issues

• We would need to ensure that records are eventually written, by having some timeout, depending on the preCommitaction
• some of the documented behaviour WRT to batching would change
• file overwrite needs to be handled specially (that isn't in the PR below, but is simple) - enough.g. they you have a filename based on key rather than unique files such as partition and offset

A simple implementation of this is in #319, but this also has some other changes