MAINT: Move to Python 3.12

[larsoner]

An auto rebuild triggered for OpenMEEG in conda-forge/openmeeg-feedstock#46 and I figured we might as well get this on our radar. Locally it still needs matplotlib so doesn't get very far:

$ ./tools/build_local.sh 
...
Channels:
 - conda-forge/label/python_rc
 - conda-forge
Platform: osx-arm64
Collecting package metadata (repodata.json): done
Solving environment: failed
...
Could not solve for environment specs
The following packages are incompatible
├─ python 3.12.0rc3**  is installable and it requires
│  └─ python_abi 3.12.* *_cp312, which can be installed;
└─ seaborn 0.12.2**  is not installable because it requires
   └─ seaborn-base >=0.12.2,<0.12.3.0a0 , which requires
      └─ matplotlib-base >=3.1,!=3.6.1  but there are no viable options
         ├─ matplotlib-base [3.3.2|3.3.3|...|3.7.3] would require
         │  └─ python_abi 3.8.* *_cp38, which conflicts with any installable versions previously reported;
         ├─ matplotlib-base [3.3.2|3.3.3|...|3.8.0] would require
         │  └─ python_abi 3.9.* *_cp39, which conflicts with any installable versions previously reported;
         ├─ matplotlib-base [3.4.3|3.5.0|...|3.8.0] would require
         │  └─ python_abi 3.10.* *_cp310, which conflicts with any installable versions previously reported;
         └─ matplotlib-base [3.6.2|3.6.3|...|3.8.0] would require
            └─ python_abi 3.11.* *_cp311, which conflicts with any installable versions previously reported.

[larsoner]

Still waiting on Numba and dipy (at least) here

[hoechenberger]

@larsoner I see you removed TensorFlow, was this necessary to get the package to build? TF is an optional, yet important dependency of SleepECG. I believe classification without TF is either impossible or extremely slow. @cbrnr at least told me something along those lines, if I recall correctly

Edit: looking at the SleepECG docs, classification is not possible w/o TF. We should try to include the TF package if possible!

[hoechenberger]

Do you think we could resolve this somehow?

Also I started transitioning away from mamba again in my personal dev environments, since conda now uses the libmamba solver by default, and it has this new feature where it can use your system certificates for TLS connections (you don't need to set a conda-specific env var anymore if you use a self-signed cert that's mandated by a corporate firewall); mamba is still lacking this feature

… this just to suggest that we may consider simply ditching mamba if it causes trouble!

[cbrnr]

Yes, SleepECG needs TF for classification.

[larsoner]

TF is not available on Python 3.12. So we'll need to revert this if it's essential