fix: Prevent panic in turbo watch with persistent tasks (#12459)
## Summary
Fixes #12458
- `turbo watch` panics at `visitor/mod.rs:564` because `Arc::into_inner`
requires the strong reference count to be exactly 1, but fire-and-forget
persistent tasks (spawned at line 491) still hold `Arc` clones of the
shared error collector when `visit()` returns.
- Uses `Arc::try_unwrap` with a fallback that locks the mutex and drains
collected errors when outstanding references remain, which is the
expected case in watch mode with persistent tasks.
## How it works
In watch mode, persistent tasks (e.g. `vite dev`) are intentionally
spawned as fire-and-forget -- the `JoinHandle` is dropped immediately so
dependent tasks can proceed. Each of these detached spawns holds a
`TaskErrorCollectorWrapper(Arc<Mutex<Vec<TaskError>>>)`. Since
persistent tasks run indefinitely, these `Arc` clones are never dropped
before the end of `visit()`.
The previous code assumed all `Arc` clones would be gone by the time it
called `Arc::into_inner`, which is only true for the non-watch code path
where every task is tracked in `FuturesUnordered` and drained.
