fix(api_call): protect node rotation and health state with a mutex

[erimicel]

Problem

Typesense::ApiCall keeps its rotation cursor (@current_node_index) and per-node health metadata (node[:is_healthy], node[:last_access_timestamp]) as plain unsynchronised instance state. When several threads share a single Typesense::Client — the standard pattern in Puma, Sidekiq, etc. — two concurrent calls to next_node can compute the same incremented index from the same starting value, causing the rotation to skip a node. A failing request can also flip is_healthy to false while another thread is mid-way through reading the pair, observing a stale last_access_timestamp with the new health flag.

The race is small in MRI thanks to the GVL, but it is real and grows under load. With keep-alive added recently, more state flows through the same path, so it's worth tightening up.

Fix

Add a single per-instance Mutex and use it to:

• protect the round-robin loop in next_node (only the cursor advance + the healthy/due check) so only one thread mutates @current_node_index at a time;
• atomically pair the is_healthy + last_access_timestamp writes in set_node_healthcheck.

The lock is acquired briefly — no I/O inside it — so contention is negligible.

Single-node deployments

The fix is safe and beneficial for single-node setups too: the rotation race is degenerate there ((idx + 1) % 1 == 0), but the health-state pair still benefits from atomic writes when a request fails and sets is_healthy: false.

Tests

Adds two regression specs in spec/typesense/api_call_spec.rb:

• multi-node: 16 threads × 90 iterations across 3 nodes; asserts each node is selected exactly 480 times.
• single-node: 8 threads concurrently flipping is_healthy and calling next_node; asserts the final state is internally consistent.

bundle exec rspec spec/typesense/api_call_spec.rb — 44 examples, 0 failures.
bundle exec rubocop lib/typesense/api_call.rb spec/typesense/api_call_spec.rb — clean.

PR Checklist

• I have read and signed the Contributor License Agreement.

[tharropoulos]

Test coverage looks good for the regression intent, though the exact even distribution is proving serialized round-robin behavior more than general thread safety.