feat(gateway): emit token_usage events after each API call for real-time context bar

[OmarB97]

Problem

The desktop app's context window progress bar only updates at the end of a turn, causing a jarring jump. Token count data is only sent to the UI via session snapshots that fire after the turn completes.

Solution

After every API call that updates session token counters in conversation_loop.py, emit a structured token.usage event through the existing status_callback → _status_update → WebSocket event pipe. The desktop app's @nous-research/ui context bar component can then subscribe to token.usage events and update smoothly during a turn.

Changes

File	Change
run_agent.py	Add _emit_token_usage() method — JSON-encodes token counts + context dimensions, fires via status_callback
agent/conversation_loop.py	Call _emit_token_usage() after session token counters are updated
tui_gateway/server.py	Handle kind=token_usage in _status_update — parse JSON payload, re-emit as typed token.usage event

Event Payload

{
  "input_tokens": 12345,
  "output_tokens": 678,
  "total_tokens": 13023,
  "context_tokens": 54321,
  "context_length": 131072,
  "context_pct": 41.5
}

Frontend Note

The @nous-research/ui npm package (v0.18.2) renders the context bar component. A separate change in that package is needed to subscribe to token.usage events. This PR provides the backend event pipe — the frontend half is independent.

Testing

• Fire-and-forget via status_callback — never blocks the conversation loop
• Graceful degradation: if no status_callback is wired, the event is silently dropped
• JSON parse errors in the gateway are swallowed
• Token counters still updated via the existing path (no regression)

[OmarB97]

Superseded by #39370, which carries the full token-usage pipeline (_emit_token_usage, in-loop emission incl. preflight estimates, gateway token.usage bridge) rebased on current main; the bundled TUI fallback-provider fix is tracked in #36056. Closing to reduce reviewer surface.