Allows to read Single protobuf message at once (w/o length-delemeters).

* Add a test for canceling distributed queries on error

* Cancel other processors on exception in PipelineExecutor

Fixes: 01514_distributed_cancel_query_on_error

Clean the aggregates pools (Arena's objects) between flushes, this will
reduce memory usage significantly (since Arena is not intended for
memory reuse in the already full Chunk's)

Before this patch you cannot run SELECT FROM huge_table GROUP BY
primary_key SETTINGS optimize_aggregation_in_order=1 (and the whole
point of optimize_aggregation_in_order got lost), while after, this
should be possible.

This will fix possible UAF, since ThreadFromGlobalPool cannot wait until
the job will be reseted, for this you need to call ThreadPool::wait()
(that waits until scheduled_jobs will be zero), and this guarantee that
the job was reseted, otherwise some variables that job is referencing
may be already destroyed, while job hasn't been destoyed yet (example
that I found PipelineExecutor -> ThreadGroupStatusPtr -> MemoryTracker
-> ~MemoryTracker -> log, while log had been destroyed already, here is
TSAN report --
https://gist.github.com/azat/d480dc3af5a0a44de4b038e20807c4b9, copied
from
https://clickhouse-test-reports.s3.yandex.net/15035/79133a426fdf042e383ea0cdccc4dc8273baa3a7/functional_stateless_tests_(thread)/test_run.txt.out.log)