fix(env): use conda-forge pytorch-gpu for reliable CUDA builds (#11)

Replace pytorch/nvidia channels + pytorch-cuda=12.4 with conda-forge
only + pytorch-gpu meta-package. This matches the working ESA-Datalabs
setup and ensures the CUDA variant is installed during Docker builds
(with CONDA_OVERRIDE_CUDA="12").

The pytorch-cuda package from the pytorch channel does not constrain
the pytorch package to its GPU variant, resulting in CPU-only builds.
The conda-forge pytorch-gpu meta-package correctly forces the CUDA
variant of pytorch.

Author: gomezzz

environment.yml

@@ -6,8 +6,6 @@
 #   the terms contained in the file 'LICENCE.txt'.
 name: am
 channels:
-  - pytorch
-  - nvidia
   - conda-forge
 dependencies:
   - astropy
@@ -23,7 +21,7 @@ dependencies:
   - pyarrow
   - python=3.11
   - pytorch>=2.6
-  - pytorch-cuda=12.4
+  - pytorch-gpu
   - pyturbojpeg
   - scikit-image
   - scikit-learn