feat(pubsub): inline Pub/Sub APIs for exact channels

[Aryex]

Summary

This adds the part 2 of PubSub APIs implementation plan.

Issue link

Related #135

Changes

• PubSub implemented to support the APIs, including the ffi callback.
• RESP3 is enforced client side.
• unit and integration tests added.

Limitations

• publish takes the message first — publish(message, channel) — matching the other GLIDE clients rather than redis-rb's publish(channel, message).
• timeout: is documented and treated as milliseconds, not seconds. Worth confirming that's the intended public unit, since the rest of the client's timeouts are in seconds.

Testing

• bundle exec rubocop — clean.
• bundle exec rake test:unit — 227 tests, 480 assertions, 0 failures, 0 errors, 0 skips. test/unit/glide/pubsub_test.rb grows by ~234 lines covering the RESP3 guard, timeout_argument conversion and rejection, and the push handler's kind filtering / NUL-safe reads / exception swallowing.
• test/integration/valkey/pubsub_test.rb — new, ~277 lines, wired into the standalone suite. Covers subscribe/publish/receive round trips, blocking vs non-blocking consume, and the negative paths.
• Integration suite has not been run against a live server in this pass; that's the main gap before this leaves draft.

Checklist

Before submitting the PR make sure the following are checked:

• This Pull Request is related to an issue.
• Commit message describe your changes
• Commits are signed off (git commit -s) per the DCO.
• Tests are added or updated.
• CHANGELOG.md and documentation files are updated.
• Linters have been run (bundle exec rubocop) and pass.
• Destination branch is correct - main.

[jamesx-improving]

CHANGELOG.md still needs a 1.1.0 entry for this — it's the largest user-facing change in the release and the checklist box is unchecked.

On your open question: milliseconds is the right wire unit — glide-core parses the last argument as a u64 ms count and Go's peer API names the parameter timeoutMs — but timeout: reads as seconds next to every other timeout in this client, so consider renaming the kwarg timeout_ms: while the API is still unreleased.

[jamesx-improving]

This also mixes PubSubCommands into Valkey::Pipeline (pipeline.rb:7 includes Commands), which never assigns @pubsub — so r.pipelined { |p| p.publish("m", "c") } raises NoMethodError: undefined method 'publish' for nil and aborts the whole batch, where PUBLISH would otherwise pipeline fine.

Suggested fix: have PubSubCommands#publish call send_command(RequestType::PUBLISH, [channel.to_s, message.to_s]) on self (keeping the sharded: guard in the mixin) so it works for both the client and a pipeline; the subscribe/consume methods aren't pipelineable and are worth an explicit Pipeline override that raises something readable.

[Aryex]

This is a good point and raised an oversight in my design of the PubSub class. In order to support PubSub in Batch, we will need to refactor PubSub into a mixin. I will handle this in a follow up PR once this is merged.

[Aryex]

I guess it is less of an in compatibility and more that it is an anti-pattern with how the current client work. See my comment.

My suggestions is to perform this refactor after this is merged. Then proceed with support for other APIs

[jamesx-improving]

timeout: nil → "0", and glide-core reads 0 as no deadline: wait_for_sync sets deadline = None and falls through to a bare notified.await, and the pubsub interception bypasses request_timeout. Since command is blocking: true, a subscribe that never reconciles (server restarted right after connect) parks the thread permanently — Timeout.timeout and Thread#raise can't interrupt a thread inside a blocking FFI call — and this PR removed the "nil blocks indefinitely" line from every public doc comment.

Suggested fix: restore that wording on the blocking methods, or default timeout: to the client's request_timeout so the failure is a Valkey::TimeoutError instead of a hang.

[Aryex]

Fixed in 730dfaa

[jamesx-improving]

"Every attach_function" isn't accurate — 16 of the 21 in bindings.rb still default to blocking: false, including create_client and create_client_from_uri (bindings.rb:238, :245), the two that install this Pub/Sub callback and that hold the GVL for the entire connect-and-retry window.

Suggested fix: add blocking: true to both create_client* functions, and scope the rule to calls that can actually block (connect, command, batch, script, close) — the free_* and OTel span helpers don't need it.

[Aryex]

Good point. updated in 587e665

I have also raised #309 to fix the create_client methods.