Feature request: batched & multi-GPU inference support for SAM3 Agent

### Feature request: batched + multi-GPU inference for SAM3 Agent

Hi, thanks for releasing SAM 3 and the Agent pipeline!

I’m currently using the Agent interface (e.g. `sam3/agent/inference.py` + `run_single_image_inference`) with a vLLM backend to run text-driven segmentation on a large image dataset on a multi-GPU server.

It would be very helpful to have:

1. **Batched Agent inference (single GPU)**
- A function like `run_batch_inference(...)` that takes a list of `(image_path, text_prompt)` pairs and runs the Agent in batch, reusing the same SAM3 model and LLM client.
- Output format could stay the same as the current per-image `*_pred.json` files.

2. **Multi-GPU inference recipe for the Agent**
- An example script (e.g. `examples/agent_multi_gpu.py`) using `torchrun` or similar to:
- Split a list of images across multiple GPUs.
- Run the Agent on each subset.
- Avoid collisions in `sam_out/` and `*_pred.json` outputs.

This would make it much easier to scale Agent-based experiments to 10k+ images on machines with 4–8 GPUs.

**Environment (example)**
- Linux, Python 3.x, PyTorch 2.x
- GPUs: multi-GPU server (e.g. 8 × A800)
- SAM3: latest from GitHub
- LLM backend: vLLM + Qwen/Qwen3-VL-32B-Instruct

If there is already a recommended way to do batched or multi-GPU Agent inference, pointers would be greatly appreciated. Thanks!

Start with sam3/agent/inference.py and run_single_image_inference to understand the current per-image Agent flow and output files. Then review the requested examples/agent_multi_gpu.py approach using torchrun, including image partitioning and output collision avoidance. Done means batched inference and a multi-GPU recipe work for the described image and prompt inputs while preserving the existing prediction output format.

Written by the indexing model from the issue text.