Add Dockerfile for ARM64/DGX Spark support

[LalithShiyam]

Summary

This PR adds ARM64 support for running the nnInteractive server on NVIDIA DGX Spark systems (and other ARM64 platforms with NVIDIA GPUs).

Key changes:

• Added Dockerfile.arm64 using NVIDIA NGC PyTorch 25.01 container as base
• Uses NGC PyTorch which is properly compiled for ARM64 with CUDA support
• Preserves NGC PyTorch during dependency installation (pip would otherwise install CPU-only version)
• Pins numpy<2 for compatibility with NGC PyTorch (compiled against NumPy 1.x)

Tested on:

• NVIDIA DGX Spark with GB10 GPU
• CUDA 12.8 / Driver 580.95

Build & Run:

cd server
docker build -f Dockerfile.arm64 -t nninteractive-slicer-server:arm64 .
docker run --gpus all -p 1527:1527 --ipc=host nninteractive-slicer-server:arm64

Test plan

• Built image on DGX Spark (ARM64)
• Verified PyTorch CUDA availability
• Server starts and accepts connections from 3D Slicer

🤖 Generated with Claude Code

[LalithShiyam]

Hi @coendevente, thanks a lot for the great work on this project.

We are using DGX Spark as our primary platform for running SlicerNNInteractive. Since the hardware is very recent, we ran into a few expected compatibility issues (e.g., CUDA 13.0, sm_121). Recent PyTorch releases and updated NGC base images resolve most of these problems.

This PR adds a new Dockerfile that installs and runs SlicerNNInteractive cleanly on DGX systems. The goal is to provide a reproducible setup for DGX-class hardware without requiring manual workarounds.

Happy to adjust or iterate if you prefer a different approach.