Support memcpy between iGPU and dGPU.

### Is your feature request related to a problem? Please describe

In llama.cpp for SYCL backend, current solution of memory copying between iGPU and dGPU is device-host-device.
With the iGPU (built-in GPU) become powerful, it's possible to run the LLM on iGPU and dGPU in parallel for different functions.

Currently, the iGPU and dGPU are different platform type and do not support memory copy between them.
We need the feature to support memcpy between iGPU and dGPU directly.

In other area, iGPU cooperate with dGPU would have more usage.
The communication between iGPU and dGPU become more important and the low efficiency of mem-copy would become the bottle-neck.

Thank you!

### Describe the solution you would like

_No response_

### Describe alternatives you have considered

_No response_

### Additional context

_No response_