feat: show context window usage beside the send button

[iuga]

Motivation

When chatting with a local model, there's no visibility into how much of the model's context window a conversation has consumed. Once the window fills up, response quality degrades and older context gets dropped silently. This meter lets users see how close they are so they can start a fresh session at the right time.

Description

Adds a small meter to the left of the send/stop button showing x% - used / total (e.g. 24% - 12.3K / 300K). Token counts update after each turn; the window size is read from the model itself. A new Token usage stats toggle in Chat Experience controls it and defaults to on.

Changes

• Provider: expose the model's effective context window from Ollama's /api/show: the modelfile's num_ctx when set (what Ollama actually truncates at), falling back to the architecture maximum (ModelDetails.contextLength).
• Agent: read per-turn token usage from the agent's checkpointed state (one state read per interrupt cycle, shared with the ask-user check) and surface it via an onUsage handler.
• Plugin: resolve the context window on load and on every settings save; if the host was unreachable at load, retry on the next successful health check; push updates to open views.
• View + styles: render the meter, reset the counter on new/loaded sessions, turn it red at >=90%. Counts formatted with Intl.NumberFormat compact notation.
• Settings: add the showTokenStats toggle (default true).

TradeOffs

• Token usage is read from the agent's checkpointed state after each turn rather than from the live stream. LangGraph's messages stream mode drops the final usage-bearing chunk, so the stream never sees token counts. Reading from state is reliable but means the meter updates once per turn (on completion), not token-by-token — an acceptable tradeoff for accuracy.
• The context window is re-resolved on every settings save instead of caching per host/model. It's one cheap local /api/show call per user-initiated save, which keeps the plumbing much simpler and self-heals stale values.