Put all calibration resources into resource manager. This fixes a race condition

triggered in int8 mode: during calibration, when two sess.run()s fetch two
different sets of ops and both of them trigger the computation of the
TRTEngineOp, two different TRTEngineOp will be created, and dev_tensors_ will
only be allocated once by the first op by the device's resource manager via
LookupOrCreate(). As a result, the second TRTEngineOp will have a empty
dev_tensors_ vector which will result in std::out_of_range error when the op
tries to access it later.

We should consider marking TRTEngineOp as stateful, the problem is it uses TF
functions which has to be stateless, but we can use function library as key to
cache them.

PiperOrigin-RevId: 228816949