[tf.data] Reimplement `tf.contrib.data.get_single_element()` as an async op.

The current `ToSingleElementOp` kernel has a synchronous
implementation, and yet it can block an inter-op threadpool thread (in
`iterator->GetNext()`) while depending on another (e.g. if the
iterator calls a TensorFlow function to produce an element). This can
lead to deadlock if the number of inter-op threadpool threads is less
than or equal to the number of concurrent activations of that
kernel. This change fixes that deadlock by moving the blocking
computation onto a background thread.

PiperOrigin-RevId: 179067816