Fix copy of tensors for TRT on iGPU (triton-inference-server#3067)

* Add logging

* Leverage zero copy support for output tensor
- When possible

* Wait for inference completion on iGPU to correctly copy output tensors

* Add missing quote in error msg

* Remove debug logging

* Fix synchronization bug

* Remove extra logging

* Use cudaLaunchHostFunc

* Re-use zero copy, make class variable

* Fix sync

* review edits

* Avoid invalidation of vector element address

* fix sync and lifecycle of payload

* Use struct to store params

* cleanup

* Update comments

* use resize instead of reserve

* direct cuda host launch in copybuffer

* Update copyright

* Remove extra cudaStreamWaitEvent

* Use memcpy on stream for inputs

* review edits

* add temp logging

* io_binding_info was already set earlier

* additional review edits

* Remove redundant block of code

* Fix logging after SupportsIntegratedZeroCopy()