Add configurable timeout for LLM provider and session summarization

[jsykes]

Replaces hardcoded 120s timeout with a timeout field in agents.defaults config, allowing users running local/slow models to avoid request timeouts.

[Leeaandrob]

@Zepan Adds configurable timeout for LLM provider calls and session summarization. Prevents hanging requests from blocking the agent indefinitely.

Recommendation: Merge. +17/-5, small and practical. Especially important for unreliable network conditions on embedded devices.

[nikolasdehor]

Clean, focused, minimal — exactly what this needs. LGTM.

The change makes the 120s timeout configurable via agents.defaults.timeout (JSON) and PICOCLAW_AGENTS_DEFAULTS_TIMEOUT (env var), with a safe fallback to 120s when unset or <= 0. It applies to both the HTTP provider client and the session summarization context.

Strengths:

• Single concern, 17-line diff
• Consistent with existing config patterns (max_tokens, temperature, max_tool_iterations)
• Example config updated
• Backward compatible — existing deployments get the same 120s default

This directly fixes #430 and #312 for users running local models (Ollama, vLLM) that need longer inference times.

Minor nit (non-blocking): no tests added, but the change is simple enough that the risk is low. If desired, a quick test verifying NewHTTPProvider with timeout=0 falls back to 120s would be nice.

I also reviewed PR #432 which tackles the same problem but bundles 4 unrelated changes — I'd recommend merging this PR first and having #432 split and rebase on top.

[nikolasdehor]

Related: #432 (PR increasing Ollama timeout) and #605 (issue requesting timeout control). All three address the same need for configurable LLM timeouts. Suggesting consolidation — this issue (#278) is the earliest and most complete description of the requirement.