feat(ui): show context-usage pie chart on the clear-context button

[sufan721]

Add a hover pie chart next to the clear-context button in both the Electron desktop and the web console, breaking the live context window down into system prompt / tool definitions / conversation / free.

What does this PR do?

Clearing context used to be a blind operation — users couldn't see how much of the context window was occupied, or by what. This PR adds a hover pie chart to the clear-context button in both the Electron desktop app and the web console, breaking the live context down into four slices:

• System prompt / Tool definitions / Conversation / Free

The donut center shows the used percentage, with per-slice token counts and a used X / Y footer.

The backend had no context-usage endpoint at all (token estimation existed but was never surfaced — Agent.last_usage is dead code), so this also adds a read-only endpoint:

• GET /api/sessions/{id}/context_usage (SessionContextUsageHandler)
• Agent.get_context_usage() — reuses the existing _estimate_* estimators; the input budget mirrors _trim_messages
• AgentBridge.peek_agent() — a side-effect-free lookup of the live instance, so hovering never spins up MCP/skills; returns available: false when the session has no live context
• build_tools_schema() extracted from the request path for reuse by the estimator (no behavior change)

Frontend (no chart library added — hand-rolled SVG donut):

• Desktop: new ContextUsageDonut + ContextUsagePopover (body portal), wired into ChatInput; getContextUsage added to the API client
• Web console: installContextUsagePopover(), old data-tooltip removed from the button so the two don't overlap, i18n keys for zh / zh-Hant / en

Type of change

• Bug fix
• New feature
• Docs
• Refactor / chore

Checklist

• I have read the Contributing Guide
• I tested this change locally
• Code comments and docs are in English
• Linked related issue (if any): closes #

[zhayujie]

Really like this — surfacing context usage on the clear button is exactly the right place for it, and the implementation is careful: reusing the existing _estimate_* estimators, peek_agent() being side-effect-free so hovering never spins up MCP/skills, marking the payload estimated, and documenting why used can exceed limit. The hand-rolled SVG donut with no extra dependency is a nice touch too.

A few notes before merge:

1. Rebase needed — it's currently CONFLICTING. Conflicts are small and desktop-only: desktop/src/renderer/src/api/client.ts and desktop/src/renderer/src/components/ChatInput.tsx. The Python backend and web console merge cleanly. Just re-apply your getContextUsage client method and the ChatInput wiring on top of current master.

2. Please drop the accidentally committed editor file — the diff adds .obsidian/plugins/obsidian-ai-translate (a gitlink/submodule). It should be removed from the PR (and consider adding .obsidian/ to .gitignore).

3. The tools slice counts all available tools' schema, but at run time _select_tools_for_injection() may inject a subset (e.g. on-demand MCP tool retrieval). Worth a one-line comment that the tools slice is an upper bound when retrieval is on, so the chart reading isn't confusing.

Overall solid work — happy to see it land once rebased.