fix: silence shutdown noise from MqWatcher and Web3 provider

[odesenfans]

Summary

Two small follow-ups to the earlier shutdown-cleanup PRs (#1139 - #1145) that eliminate two remaining classes of noisy log output observed when Ctrl-C'ing pyaleph during a fresh sync.

1. MqWatcher prefetched-message warnings

Every Ctrl-C produced a flurry of:

[WARNING] aio_pika.queue: Message IncomingMessage:{...} lost for consumer with no_ack <RobustQueueIterator: ...>

MqWatcher._check_for_message opened the queue iterator with no_ack=True even though it only uses messages as level-triggered signals (the actual data lives in pending_txs / pending_messages). aio_pika unconditionally logs a warning per prefetched message still in the iterator buffer at close time when no_ack=True.

Switched to no_ack=False with explicit await message.ack(). Buffered messages are now nack'd with requeue=True on shutdown, so nothing is lost and the warning disappears.

2. Unclosed aiohttp session from the ETH Web3 provider

After SIGTERM:

[ERROR] asyncio: Unclosed client session
[ERROR] asyncio: Unclosed connector

web3.AsyncHTTPProvider caches an aiohttp.ClientSession per endpoint in its HTTPSessionManager. The provider exposes disconnect() to close it, but the only call site was inside EthereumConnector.packer's finally. Reader-only nodes (no packing_node) never reach it, so the session leaks at shutdown.

• EthereumConnector is now an async context manager whose __aexit__ calls provider.disconnect().
• ChainConnector is also an async context manager. It owns an internal AsyncExitStack and enters every chain connector that supports the protocol, dedup'd by id() since EthereumConnector is registered as both reader and writer.
• commands.py registers chain_connector with the main AsyncExitStack, alongside the other shutdown callbacks added in the previous PRs.
• The duplicate provider.disconnect() in packer's finally is removed (cleanup now lives on the context manager).

Test plan

• hatch run testing:test tests/chains tests/message_processing (104 passed, 3 skipped)
• black / isort / ruff clean on the four modified files
• Local smoke test: just start-dev-env && just run, wait for ETH messages to arrive, Ctrl-C, confirm the two warning families are gone

[foxpatch-aleph]

Clean follow-up PR that eliminates two sources of shutdown noise: MqWatcher warning from aio_pika on prefetched unacked messages, and unclosed aiohttp session from the Web3 provider. Both fixes are correct and minimal. The ChainConnector async context manager properly deduplicates by identity and skips connectors without cleanup needs. No logic errors or regressions found.

src/aleph/chains/connector.py (line 50): The id() dedup is safe here since connectors are long-lived singletons, but worth noting that this relies on object identity stability. If a connector were ever recreated with a new identity but the same logical instance, the dedup could break. Consider using the connector object itself as a dict key instead of id(connector) for slightly more robustness, though this is a nit.

[foxpatch-aleph]

Well-targeted PR that fixes two sources of noisy shutdown warnings. The MqWatcher change (no_ack=False with explicit ack) correctly avoids aio_pika's 'message lost' warning while preserving message delivery via requeue. The EthereumConnector/ChainConnector async context manager refactor ensures the aiohttp ClientSession is properly closed on shutdown even for reader-only nodes, removing the 'Unclosed client session' error. The deduplication by id() in ChainConnector.aenter correctly handles connectors registered as both reader and writer. The removal of the redundant provider.disconnect() from packer's finally block is a good cleanup since the context manager handles it now. All changes are well-reasoned and correctly implemented.