streaming: update the stream ingestion processor to use a flushing buffer

[pbardea]

Currently the stream ingestion processor uses a slice as it's buffer before it flushes. We should have some sort of BufferingAdder that supports inserting keys at a particular timestamp. Concrete actions here:

• Consider augmenting the BufferingAdder interface to support inserting MVCC keys.
• Add support for a kvBuf type buffer that supports storing MVCC KVs.

Jira issue: CRDB-3325

Epic CRDB-19048

[blathers-crl[bot]]

cc @cockroachdb/cdc

[stevendanna]

I think we still want to do this. At the very least, we need memory monitoring around the buffer in the ingestion processor.

[gh-casper]

Put some time on reusing the buffer_adder in stream_ingestion_processor, it can end up very messy code. After some research on this, I think we better have its own buffer_adder, sst batcher and stats for stream ingestion:

1. better observability on flushes, latency, memory monitoring.
2. consolidate current range key batcher with the mvcc key batcher into one piece.
3. more control on how we flush and split the sst we are about to ingest.

However, we can still reuse the kvBuf for the way it encode keys and values, it needs some modifications mostly how we sort it kvBufEntry if we want to encode the key and timestamp into one fat key.