Add support for gpu selection for users with multiple gpus

[RngDevil]

Purpose

For users that have more than one GPU (e.g. Laptop with RTX 4060 8GB - gpu-id 0 and eGPU-RTX 5060 Ti 16GB - gpu-id 1) and would like to select the more performant GPU (e.g. gpu-id 1)

Changes

1. Add optional command line parameter --gpu-id <GPU_ID> to select the GPU. Defaults to 0 for backwards compatibility
2. Changes hardcoding of device_id=0 to device_type=device_type and device id=gpu_id where appropriate
3. Tested on 4060 and 5060 using TensorRT

[RngDevil]

Note, when changing to different GPUs, the cached tensorrt-engines may need to be deleted or at least renamed

[Glat0s]

Review against current upstream/dev for merge safety:

This PR is not safe to merge into dev as-is. Please repost to dev branch.

1. Merge conflict against dev

   The PR is currently based on main, and GitHub reports it clean for that base, but merging the PR head into current upstream/dev conflicts in:

   • app/processors/workers/frame_worker.py

   Repro:

   git fetch upstream dev pull/231/head:review-pr-231
   git merge-tree upstream/dev review-pr-231

   Result includes:

   CONFLICT (content): Merge conflict in app/processors/workers/frame_worker.py
   

   Fix: rebase/retarget this PR onto the latest dev branch and resolve the frame_worker.py conflict, then rerun the processor tests.

2. CPU provider can receive an invalid I/O binding device id when --gpu-id is nonzero

   ModelsProcessor.switch_providers_priority() keeps self.gpu_id unchanged when switching to CPU, while many ONNX Runtime I/O bindings now pass device_type=self.models_processor.device_type together with device_id=self.models_processor.gpu_id. Example paths include app/processors/face_masks.py where CPU mode would bind device_type=cpu and device_id=<gpu id>.

   That is risky because CPU I/O bindings should use device id 0; running with --gpu-id 1 and then selecting CPU can pass device_id=1 to CPU bindings.

   Fix: centralize the binding device id, for example:

   @property
   def binding_device_id(self) -> int:
       return self.gpu_id if self.device_type != cpu else 0

   Then use self.models_processor.binding_device_id for all ORT bind_input() / bind_output() calls. Alternatively, update gpu_id-based binding sites to use 0 whenever device_type == cpu.

Validation performed:

• Targeted PR Add support for gpu selection for users with multiple gpus #231 processor tests excluding the baseline-crashing faceutil test: 56 passed.
• tests/unit/utils/test_faceutil_math.py aborts in this local environment on both PR Add support for gpu selection for users with multiple gpus #231 and current upstream/dev, so I did not count that as a PR regression.

[RngDevil]

Changes have been rebased to dev branch, resolved the conflict and resolved the CPU provider can receive an invalid I/O binding device id when --gpu-id is nonzero issue. I've also tested this locally on my system to ensure it is still functional

[Glat0s]

Would be nice to have the gpu selection options in the ui also. We can do this later. I'll merge for now.