Gemini agentic tool runs were failing with opaque "max_tokens limit"
errors when the configured max output tokens were too small for
multi-step tool calls. Lift Gemini's effective output budget to a
practical floor and surface clear guidance in both the core agent
and the QGIS chat dock when a model output cap is hit.

- Streaming path now treats `stop_reason=max_tokens` as a failure and routes
  the error through `_format_chat_worker_error`, so users see the same
  actionable max-token guidance during streamed runs instead of a truncated
  answer.
- Replace the duplicated max-token marker list in `_format_chat_worker_error`
  with the shared `_looks_like_max_tokens_reached` helper from
  `geoagent.core.agent`, keeping the QGIS UI and core agent in sync (now
  also catches `stop reason: max_tokens`).
- Add a test covering both `stop_reason=max_tokens` and `stop reason: max_tokens`
  phrasings so the two paths cannot drift again.

…e configuration

- Clamp positive max-token values below 256 up to 256 in
  `_max_tokens_to_setting`. The settings spinbox now uses 0 as the Auto
  sentinel, which lets users type 1-255 that the previous min=256 spinbox
  rejected. Saving normalizes those into the historical 256-token floor so
  the saved config never holds a value the spinbox previously disallowed.
- Add a unit test that pins the clamp behavior and the Auto sentinel.