[tf.data] Fixes for two recently introduced use-after-free bugs.

1. In ParallelMapIterator, do not call `cond_var_.notify_all()` without holding
   the associated mutex. In some cases, the iterator may have been deleted
   between releasing the lock and notifying the condition variable, which
   leads to a use-after-free. This change applies this style to all use of
   condition variables in tensorflow/core/kernels/data/.

2. In CapturedFunction::RunAsync(), do not use `shared_ptr` to manage
   the lifetime of objects that (potentially) borrow from runtime
   objects. The present code runs the destructor after the `done()`
   callback is called, but the `done()` callback may be the last
   action in a session, and thus trigger destruction of those borrowed
   objects. In that case, the `shared_ptr` destructor may use the
   borrowed objects after they are freed.

PiperOrigin-RevId: 213872829