[TF:XLA] Simplify implementation of XlaAssignVariableOp.

Remove XlaDeviceContext::CopyDeviceTensorToDevice because it is now unused.

In https://github.com/tensorflow/tensorflow/commit/ca69ddc34b37258534d8327ec55a26b2add6a632#diff-6bcd989aebc981d5f8428a6b5da4eed5 the equivalent change was made to the core AssignVariableOp; the same reasoning applies here.

(In fact, the only reason we need a different XlaAssignVariableOp kernel is because the TF core one is templated on the data type ? the implementation here is identical to one of the cases of that op.)

PiperOrigin-RevId: 218676071