dinov2: fix environment drift by pinning cuda-version=11.7 and deps

[piotr-bojanowski]

Summary

This PR updates and stabilizes the conda.yaml environment used for DINOv2 training and evaluation.
The main goal is to improve reproducibility and avoid dependency drift/runtime mismatches while keeping the PyTorch 2.0 stack.

What changed

Updated base env to Python 3.10

Kept core stack aligned to:

• pytorch==2.0.0
• pytorch-cuda==11.7
• torchvision==0.15.0

Added compatibility pins for stability:

• mkl=2024.0.0
• numpy<2
• setuptools<81
• cuda-version=11.7

Kept required runtime dependencies in env:

• xformers==0.0.18
• submitit
• cuml

Why this is needed

During reproducibility testing, we hit environment-level issues including:

• Torch import/runtime mismatch (iJIT_NotifyEvent) resolved by pinning MKL.
• CUDA meta-package drift (cuda-version resolving to an unintended line) despite pytorch-cuda=11.7, resolved by explicitly pinning cuda-version=11.7.
• torchmetrics==0.10.3 requiring pkg_resources availability, handled by pinning setuptools<81.
• NumPy ABI compatibility warnings with NumPy 2.x, handled by pinning numpy<2.

This PR tightens the critical version boundaries so train/eval workflows are reproducible.

Validation

Verified successful environment creation and imports (torch, xformers, submitit, cuml).
Verified training runs complete with the updated environment.
Verified eval entrypoints run with the updated dependency set.

[patricklabatut]

Ideally the requirements.txt would also be updated. Also suspect this update might break mmcv / mmseg support for the dense evaluations. Well, so be it...