Harden MultiprocessingWriter shutdown and error handling

[bednar]

Context

PR #217 removed the generated OpenAPI write stack and introduced the internal RestClient. PR #222 refactored MultiprocessingWriter.

The current multiprocessing implementation still has shutdown and error-handling risks that can cause inconsistent state or a blocked queue.join().

Scope

Improve MultiprocessingWriter reliability:

• Replace the assert used for runtime state validation with an explicit exception.
• Always call queue.task_done() when processing a queued write, including failure paths.
• Add an explicit, idempotent close() method.
• Use close() from the context manager and keep __del__ as best-effort cleanup only.
• Ensure on_shutdown is called at most once.
• Handle queue timeout and worker shutdown deterministically.

Tests

Add tests covering:

• writing before start();
• writing after shutdown;
• worker timeout;
• write failure without a queue.join() deadlock;
• repeated close() calls;
• callback invocation count.

Acceptance criteria

• A failed worker write cannot leave queue.join() blocked indefinitely.
• on_shutdown is invoked at most once.
• Shutdown is deterministic for both explicit close() and context-manager usage.
• Existing multiprocessing behavior remains compatible.
• All unit and integration tests pass.

Out of scope

• Removing multiprocessing support.
• Redesigning batching or retry behavior.
• Adding new InfluxDB endpoints.
• Changes to the public client API beyond the lifecycle and error-handling fixes required here.