fix: use ConnectionResetError in TLSUpgradeProto.connection_lost

[yemreck]

Summary

Fixes #1310.

TLSUpgradeProto.connection_lost() synthesizes a generic ConnectionError("unexpected connection_lost() call") when the server closes the connection during SSL negotiation (exc=None). This error type is not handled by asyncpg's own Connection._cancel(), which only catches ConnectionResetError:

# asyncpg/connection.py - Connection._cancel()
except ConnectionResetError as ex:
    # On some systems Postgres will reset the connection
    # after processing the cancellation command.

The generic ConnectionError falls through to the except (Exception, ...) handler and leaks into application code, causing confusing unhandled errors.

The problem in practice

When pool_pre_ping sends a trivial query (fetchrow(";")) that gets cancelled (e.g. due to asyncio task cancellation under load), the following cascade occurs:

1. fetchrow(";") gets CancelledError
2. The finally block tries await tr.rollback()
3. The rollback triggers Connection._cancel() via the cancel side-channel
4. _cancel() opens an SSL connection → TLSUpgradeProto → server closes during negotiation → connection_lost(exc=None)
5. ConnectionError("unexpected connection_lost() call") propagates unhandled to application code

The fix

Use ConnectionResetError instead of ConnectionError. This is:

• Semantically correct — the server reset/closed the connection during SSL negotiation
• Properly handled — Connection._cancel() already catches ConnectionResetError explicitly
• Consistent — error message follows the same format as the sibling error in data_received ('PostgreSQL server at "{host}:{port}" rejected SSL upgrade')
• Backward compatible — ConnectionResetError is a subclass of ConnectionError, so any existing except ConnectionError catches still work

Test plan

• Existing test suite passes (no tests reference "unexpected connection_lost")
• Verify ConnectionResetError is caught by Connection._cancel()'s existing handler