[muhammad-musa-ml]

Fixes #602.

Buffer passed a filtered message to the next step even while earlier messages were still waiting in the buffer, so CommitOffsets could commit past messages that had not been flushed. The issue has the Snuba and Sentry chains and the broker's commit calls.

The offsets of a filtered message are now merged into the batch and go out with it. When Buffer holds no offsets, the message is forwarded at once, as before. The check uses the offsets Buffer holds and not the buffer's is_empty, because a buffer whose compute_batch_size returns 0 reports empty while it holds a message.

Three new tests in tests/processing/strategies/test_reduce.py; all three fail on main and pass here:

• a filtered message is forwarded only when nothing is waiting, including a batch of size 0
• filtered offsets are held until the batch is flushed, with two partitions
• FilterStep -> Reduce -> RunTask -> CommitOffsets commits nothing before the batch is written

Run on Windows with Python 3.13.5 and confluent-kafka 2.16.1:

• the suite without the Kafka-backed tests (no broker here) and without the multiprocessing tests: 131 passed, 6 skipped, 1 xfailed. The multiprocessing tests give 20 failed, 2 passed and 15 errors on main and with this change.
• mypy . --strict: only the three signal.SIGCHLD errors that main also gives on Windows.
• flake8, isort and black (22.3.0 and 26.10.1) are clean on the lines I changed. black would also reformat the flush_reason line in buffer.py on main; I left it.

With BatchStep followed by UnbatchStep, the held offset of a filtered message at the end of a batch is not forwarded, because Unfold uses the batch's committable only when the last message has none. The last commit is then the last accepted message, and the filtered tail is read again after a restart.