The CLI returns per-turn usage data on each assistant message but the SDK
was dropping it during parsing. This exposes it as an optional dict on
AssistantMessage so consumers can track context token usage mid-stream.

When tasks don't respond to cancellation cleanly, anyio's
_deliver_cancellation() enters a busy-loop pegging CPU at 50-100%.
Set a 5s deadline on the task group's own cancel scope so it times
out instead of spinning indefinitely.

Ref: #378

Support the --worktree CLI flag for creating git worktrees per session.
Accepts True for auto-generated names or a string for explicit naming.

feat: add worktree option to ClaudeAgentOptions

When Query.close() was called from a different asyncio task than the
one that called Query.start(), anyio's CancelScope.__exit__ raised
RuntimeError ("different task") and never completed. This left
_deliver_cancellation in an infinite call_soon loop, burning ~4% CPU
per leaked scope.

The root cause was start() calling __aenter__() on the task group in
one task and close() calling __aexit__() from another. anyio requires
these to happen in the same task.

Fix: move the task group into a dedicated _run_reader coroutine
launched via asyncio.create_task, so enter and exit always happen in
the same task. close() now closes the transport first (letting
_read_messages exit naturally via sentinel/EOF), then awaits the
reader task's completion signal with a timeout fallback.