[FEAT]: Migrate AI model inference to GPU acceleration

[marcvergees]

📝 Description

Migrate the AI model inference pipeline to use GPU acceleration so that model inference runs significantly faster and more efficiently. This includes enabling CUDA/CUDNN support where applicable and switching to GPU-enabled builds of the inference libraries.

💡 Rationale

Current CPU-based inference is slow for larger inputs and high-throughput workloads. GPU acceleration will reduce latency and increase throughput for AI-driven features in FireForm, improving user experience and enabling more complex models to be used in production.

🛠️ Proposed Solution

A brief plan to implement GPU acceleration:

• Add GPU-capable dependencies (e.g., torch with CUDA support, CUDA-enabled runtime images) and document compatibility in requirements.
• Update inference code in src/ to select GPU when available and fall back to CPU.
• Add Dockerfile variant or runtime configuration for CUDA-enabled containers (nvidia/cuda base images or NVIDIA Container Toolkit support).
• Integrate optional TensorRT/ONNX Runtime backend for further optimization (if applicable).
• Add CI/Dev steps for smoke testing on GPU-enabled runners or document how to test locally with NVIDIA Docker.
• Add performance benchmarks and example configs for common model sizes.

✅ Acceptance Criteria

• GPU-enabled Docker image runs inference successfully.
• Code auto-detects GPU and falls back to CPU when not available.
• Documentation updated in docs/ with setup and testing steps for GPU.
• Measurable performance improvement documented (e.g., latency and throughput benchmarks) for one or more representative models.

📌 Additional Context

• Consider compatibility matrix for supported CUDA/CUDNN and PyTorch/TensorRT versions.
• If the project supports multiple model backends (Mistral/Ollama/etc.), document which backends can use GPU acceleration and any extra setup required.
• This change may require updates to deployment documentation and CI to support GPU testing or clear instructions for maintainers on how to validate changes.