feat(cli, llm): Add jp conversation title command

[JeanMertz]

Users can now regenerate a conversation's title on demand with jp conversation title (alias t), without going through conversation edit --title. It offers the generated candidates as a picker that can request another batch or fall back to a hand-written title, supports --count/-n to change how many candidates are generated, --model/ -m to run generation on a specific model, and --dry-run to print candidates without applying one. Without a terminal, the first candidate is applied automatically. conversation edit --title (bare) now delegates to this same path, so both commands behave identically.

Title generation, previously duplicated between conversation edit and TitleGeneratorTask, is consolidated into jp_llm::title::generate and resolve_model. Both callers now also fit the conversation into the title model's context window, dropping older turns the same way the inquiry backend already did; that truncation logic is extracted into a new jp_llm::window module shared by inquiry handling, title generation, and conversation summarization.

jp conversation compact --summarize now rejects a turn range that doesn't fit the summarizing model's context window instead of silently truncating it, since a stored summary has to cover the full range it claims. The error suggests narrowing --from/--to or picking a larger-window model.

Provider errors that report an oversized prompt (Anthropic, OpenAI, Google) are now classified as ContextWindowExceeded instead of falling through to generic retry heuristics, since retrying sends the same oversized request again.