fix: upgrade Docker build to CUDA 12.6 and Python 3.11

[sameeul]

Summary

Upgrades the GPU Docker build to CUDA 12.6 and Python 3.11:

• Dockerfile base image nvidia/cuda:11.5.2-runtime-ubuntu20.04 → nvidia/cuda:12.6.0-runtime-ubuntu22.04
• conda_linux_gpu.txt: replace cudatoolkit=11.5 / cudatoolkit-dev=11.5 with the CUDA 12.6 conda-forge packages that actually ship nvcc (cuda-version=12.6, cuda-nvcc=12.6, cuda-cudart-dev=12.6, libcufft-dev)
• build_conda.sh: add -DCMAKE_CUDA_COMPILER=$CONDA_PREFIX/bin/nvcc so CMake doesn't rely on check_language(CUDA) discovery, and bump PYTHON=3.9 → 3.11 (3.9 is EOL)
• publish_docker.yml: update the hardcoded nyxus-3.9 env path to nyxus-3.11

Root cause (the CUDA fix)

cudatoolkit-dev=11.5.0 installs headers/libs but doesn't place nvcc where CMake's check_language(CUDA) looks, so CMake reported CMAKE_CUDA_COMPILER=NOTFOUND and hit the fatal error in CMakeLists.txt. The CUDA 12.6 conda-forge packages ship nvcc, and passing -DCMAKE_CUDA_COMPILER explicitly makes discovery deterministic.

Rebased onto the z5 3.0.1 upgrade

This branch previously carried two extra commits that worked around the old z5 2.0.20 + xtensor stack in the Docker build:

• "check for z5/multiarray/xtensor_access.hxx instead of z5/z5.hxx"
• "install z5 2.0.20 C++ headers in Docker build script"

Both are moot now that the z5 3.0.1 upgrade is merged into main: CMake looks for z5/z5.hxx again, and the z5 C++ headers come from the z5py >=3.0.1 conda package in conda_cpp.txt (no manual header install needed). The branch has been rebased onto the current main and those two commits dropped, leaving only the Docker/CUDA/Python change.

Note the Python 3.11 bump is now also required for the GPU build to solve, not just a nicety: z5py 3.0.x has no Python 3.10/3.9 build (3.11+ only), and build_conda.sh pulls z5 via conda_cpp.txt (z5py >=3.0.1).

Test plan

• Trigger the Publish Docker image workflow on this branch and confirm the "Run Build Script" step passes (CUDA build + z5 3.0.1 via z5py on Python 3.11).

🤖 Generated with Claude Code