pytorch-gpu: Try conda instead of mamba to fix CI

This change should be temporary and can only be seen using kaniko. With
`docker build`, even with `--no-cache`, there is no issue with the
dependencies and it works with mamba. In kaniko, the dependency
resolution does not work with mamba, but with conda it does.

pytorch-gpu/Dockerfile

@@ -33,7 +33,7 @@ RUN wget --no-hsts -q -O vscode_cli.tar.gz 'https://code.visualstudio.com/sha/do
 # install some packages
 # NOTE: without --no-update-deps somehow an undefined symbol error in /opt/conda/lib/python3.11/lib-dynload/_sqlite3.cpython-311-x86_64-linux-gnu.so occurs, which crashes jupyter lab.
 USER ${NB_UID}
-RUN mamba install --yes --no-update-deps \
+RUN conda install --yes --no-update-deps \
     # umap + hdbscan
     'umap-learn' \
     'hdbscan' \