Fix: make StreamableHTTP#close and Native::Client#stop exception-safe

[raySavignone]

Problem

Client teardown isn't exception-safe in two places, so a failing session DELETE strands resources. For hosts that build short-lived StreamableHTTP clients inside a long-running process, these accumulate — leaked SSE threads, HTTPX clients, and per-client HumanInTheLoopRegistry entries (each of which owns a live scheduler thread).

1. StreamableHTTP#close runs terminate_session → cleanup_sse_resources → cleanup_connection with no ensure. terminate_session raises a TransportError on a failed session DELETE (5xx / network error), skipping the other two steps.
2. Native::Client#stop releases the registry (and resets state) only after @transport.close, so a raising close skips it.

Fix

Wrap the teardown in ensure in both methods so the remaining cleanup always runs, while the original error still propagates. This matches the gem's existing defensive pattern — close_client, close_all_clients, and HumanInTheLoopRegistry#shutdown already use rescue/ensure.

Tests

Added specs that simulate a raising session teardown and assert the SSE/HTTPX resources and the per-client registry are released, and that the original error still propagates.