ALEC-308: One LLM exchange for both features, with time for local models

[joseanesONMS]

Jira: https://opennms.atlassian.net/browse/ALEC-308

Second slice of ALEC-308, stacked on #177. Not a chat feature: /chat/completions is the name of the HTTP API every hosted or local LLM server exposes, and this is the one place ALEC now talks to it.

• The problem: root cause analysis and LLM clustering each had their own copy of the request/response code, and both gave up after 30 seconds; a local model prefilling a large alarm set and generating a grouping needs longer, so every clustering pass against LM Studio or Ollama timed out before the answer arrived and nothing was ever recorded.
• The exchange now lives once in a new features/llm-client bundle (LlmExchange): build the request, force a tool call, run any data-tool calls the caller offers (none yet; the MCP tools come in the next slice), stop at the terminal call, record token usage per round into the ALEC-308: Chart LLM token usage on the Resource Graphs #177 sink.
• Both features call it in single-shot mode, so the wire format they send is unchanged.
• Read timeouts rise to 180 s per request, 60 s for the interactive validation probe, and the pending-analysis stale window grows to match.
• A malformed terminal tool call is now a failure instead of an empty success, tokens billed before a failure stay accounted for in usage rows, budgets and gauges, and a base URL a local server rejects gets a hint to append /v1.
• Verified against LM Studio (gemma-4-e4b): a clustering pass completes under the new timeout; 30 exchange tests plus the adapted engine and suggestions suites pass and the Karaf feature verification stays green.