[HAL] Device side command buffer execute

### Request description

CUDA graph allow to launch graphs from device (*), and it allows to implement on-device while loops, while today they are forced to do a roundtrip to host to get the predicate value from device.

Looking at HAL, it doesn't look like this can be expressed in the current API. Anything like that on the vulkan/metal side? Can this be added to HAL and be backend-agnostic-enough?

(*) https://developer.nvidia.com/blog/enabling-dynamic-control-flow-in-cuda-graphs-with-device-graph-launch/

### What component(s) does this issue relate to?

Runtime