Add dedicated build config to streamline setting up CUDA-enabled Tesseracts

[andrinr]

Summary

Add a cuda: block to build_config that installs a CUDA toolkit on top of a non-CUDA base image, so Tesseract authors don't have to hand-roll the apt-keyring + component install + arch-list dance.

Motivation

CUDA-from-source Tesseracts currently carry ~10 lines of boilerplate per image:

• NVIDIA apt keyring + repo setup
• cuda-nvcc, cuda-cudart, libcublas, libcusparse component installs
• TORCH_CUDA_ARCH_LIST env var with the target SM list

This is duplicated across every CUDA-from-source Tesseract and is the slowest part of those builds (20+ min end-to-end).

Proposal

build_config:
  cuda:
    version: \"12.3\"
    components: [nvcc, cudart-dev, cublas-dev, cusparse-dev]
    arch_list: [\"7.0\", \"7.5\", \"8.0\", \"8.6\", \"9.0\"]

Behavior:

• Adds the NVIDIA apt repo + keyring for the requested CUDA version.
• Installs the requested components via apt.
• Sets TORCH_CUDA_ARCH_LIST from arch_list so PyTorch / extensions compile for the right SMs.
• Sane defaults for components and arch_list if omitted.

Impact

CUDA-from-source builds are the slowest Tesseract images in practice (~20+ min). Centralizing the install also makes it easier to bump CUDA versions across the fleet.

[jpbrodrick89]

I'm OOO today so just dipping in quickly, trying to understand what the use case is? Why can't we just recommend using a cuda base image? Or provide our own base images or would the list be too big?