[TF PyFunc] PyFunc cross-device cleanups, fix a performance bug.

* Fix overly conservative check in PyFunc devices.  This allows PyFunc to be
  used more flexibly across clusters.
* Mark PyFunc as expensive (by default if running on a GPU device, PyFunc was
  always marked as cheap, which meant it would block the main executor thread).

The second change should speed up graphs that mix PyFunc and other ops on GPU
devices by moving the PyFunc off the critical executor path
(PyFunc is registered for running on GPU devices, which is why this matters).

PiperOrigin-RevId: 236054189