Live streaming for Task API cloud agent sessions (#319077)

* Live streaming for Task API cloud agent sessions

Adds TaskTurnStreamer to incrementally push v2 Task API events into
vscode.ChatResponseStream for both initial open (mode: 'current') and
follow-up sends (mode: 'next'), mirroring the v1 JobsApiBackend SSE flow.

Key design (modelled after github-ui/agent-sessions):
- Eager tool rendering from assistant.message.toolRequests, since
  tool.execution_complete does not fire for many agent-host-synthesised
  setup ops (run_setup, run_custom_setup_step).
- Per-messageId content dedup: Task API sends progressive snapshots, not
  deltas; only the suffix is fed to the renderer.
- Intermediate assistant.message.content (content + toolRequests in the
  same event) is suppressed; only pure-text assistant messages (the final
  reply of the turn) are rendered as markdown.
- Default 'Working...' progress label until the first assistant.intent
  arrives.
- Bounded poll loop: exits on terminal state, cancellation, or
  MAX_CONSECUTIVE_FETCH_FAILURES.
- Follow-up (mode: 'next') only advances when task.sessions.length grows,
  to avoid premature exit when trailing prior-turn events arrive between
  the snapshot and sendFollowUpToTask resolving.

ChatSessionContentBuilder updates:
- buildTaskHistory: single-pass turn split + bootstrap event suppression.
- buildTaskResponseTurn: mirrors the live rules (eager tool rendering,
  intermediate-narration suppression, dedup with live cards) so the
  refresh() after streaming is idempotent.
- parseToolCallDetails: friendly cards for run_setup,
  run_custom_setup_step, report_intent.
- toFileLabel: handles v1 (/home/runner/work/...) and v2
  (/tmp/workspace/..., /workspace/...) layouts.

* Address PR review: parallelize fetches, fix stale comment, add Task API history tests

- Parallelize fetchTask/fetchEvents with Promise.all in both the main
  poll loop and _waitForTurnStart so each poll only pays the slower of
  the two requests.
- Update StreamBaseline JSDoc to reflect that phase 1 (mode: 'next')
  unblocks only on turnCount > priorTurnCount, not 'any unseen event'.
- Add unit tests covering the new Task API history rules:
  bootstrap suppression, turn-boundary split, eager tool rendering with
  execution_complete dedup, intermediate-narration suppression, and
  the synthesised-turn fallback when no user.message has arrived.

* Readme update