sql: fix auto-retry behavior with autocommit_before_ddl

[rafiss]

When the autocommit_before_ddl setting was enabled, we were
sending a notice to the client without any buffering. This prevents
auto-retry logic for retriable errors from kicking in, since if results
have already sent to the client, it's not safe to retry the current
statement.

The fix is to buffer the notice for sending instead. It will get sent
whenever results are flushed back to the client. In order to achieve
this, I introduced a new BufferClientNoticeInConnection function. The
existing BufferClientNotice function is not sufficient since that only
buffers notices in the command result, and that buffer is discarded due
to how the connExecutor state transitions are defined. Namely, the
connExecutor will execute the schema change command twice: once to
autocommit the current transaction, and another time to run the schema
change. To avoid losing the notice from the autocommit, it must be
buffered all the way into the client connection, and not just the
command result.

Running with this patch makes logic tests less flaky, since schema
changes that encounter retryable errors are way more likely to be able
to be automatically retried now. This allows us to remove the testing
knob that overrode the transaction liveness threshold. I verified the
flakiness is gone by using:

./dev testlogic ccl --config=3node-tenant --stress --ignore-cache

informs #133180
Release note (bug fix): Fixed a bug that prevented transaction retry
errors encountered during implicit transactions from being automatically
retried internally if the autocommit_before_ddl session variable was
enabled and the statement was a schema change.

[cockroach-teamcity]

This change is 

[fqazi]

Subtle bug, nice find!

Reviewed 11 of 11 files at r1, all commit messages.
Reviewable status: complete! 1 of 0 LGTMs obtained (waiting on @rafiss)