Add bulk keyword reads with capability detection, and test both transports

[mikelangmayr]

• Add keys.read, a third auto-registered meta-service that reads many keywords in one request; per-keyword failures ride inside values so one broken getter costs only itself, and pattern selection skips write-only keywords
• Report a services field from keys.list, derived from the keys this peer actually serves minus its registered keywords, so a caller can tell "no keys.read here" from "no answer"; peers on an older libby omit it, which is the negative signal
• knows_key cannot answer that question: it reads the discovery registry, which stays empty because Libby.rabbitmq hardcodes discover=False, and probing by calling keys.read is indistinguishable from a timeout since an unknown key is dropped without an ACK
• Add Client.listing, list, describe and read. list returns qualified names that feed straight back into get/show/read; listing returns names and services from the one response; read issues one request per peer, never raises for a failed read, and chunks long name lists
• Switch the CLI's list and describe onto Client, replacing its private _rpc_keys_* helpers. show <pattern> deliberately keeps one show per match, so the CLI still works against daemons on a libby without keys.read
• Parametrize tests/test_client_integration.py over ZMQ and RabbitMQ via shared base cases; the ZMQ cases need no external service and always run, the RabbitMQ ones keep their broker skip
• Drop the sleep(2.0)/sleep(1.5) warm-ups: both transports already block in start() until reachable, and test_first_read_needs_no_warmup now asserts that per transport so a barrier regression fails loudly instead of going flaky
• Add a Tests workflow, since CI built docs and never ran the suite. No broker service container, so the ZMQ cases are what guard the wire
• Document keys.read, the services field and the new Client methods in docs/source/keywords.md and docs/source/client.md
• Verified against a daemon running a pre-keys.read libby: services comes back empty and the per-keyword fallback reads correctly. 71 tests pass, pylint 10.00/10 on the new modules