fix: fall back to onnxruntime/coremltools on Python 3.12+

[bodapatisaikrishna]

Description

Fixes #203 (and the same root cause behind #159 and #188): on Python 3.12,
pip install basic-pitch fails to resolve at all, on any platform.

tensorflow/tensorflow-macos never published a cp312 wheel below version
2.16 (verified against PyPI's release metadata), but the dependency markers
cap both at <2.15.1 while still activating for python_version >= '3.11'
(non-Darwin) / python_version > '3.11' (Darwin) -- an unsatisfiable
constraint on 3.12.

Simply raising the version cap is not a safe fix. I installed
tensorflow-macos==2.16.2 and confirmed the bundled saved_models/icassp_2022/nmp
SavedModel fails to load:

AttributeError: '_UserObject' object has no attribute 'add_slot'

This is Keras 3 (TF 2.16's default) being unable to reconstruct the old
optimizer's object graph inside the SavedModel checkpoint. So relaxing the pin
would trade an install-time failure for a worse runtime one.

Fix

basic_pitch already bundles the model in three other formats
(nmp.onnx, nmp.tflite, nmp.mlpackage) and already has fallback logic in
basic_pitch/__init__.py (TF_PRESENT → CT_PRESENT → TFLITE_PRESENT →
ONNX_PRESENT) to pick whichever backend is installed. On Python 3.12+:

• Darwin: coremltools is already an unconditional dependency there, so no
  new dependency is needed -- it becomes the default backend automatically.
• Non-Darwin (Linux/Windows): added onnxruntime (has cp312+ wheels) as the
  dependency, so onnxruntime becomes the default backend.

tensorflow/tensorflow-macos (both in dependencies and the [tf] extra)
are now excluded on python_version >= '3.12' instead of installing something
that can't load the model. Added the Python :: 3.12 classifier.

Testing Instructions

Verified with real inference, not just pip install --dry-run:

1. Confirmed the vulnerability/bug is real: tensorflow-macos==2.16.2
   raises the add_slot error above when loading the bundled model.

2. Confirmed ONNX and CoreML produce matching output to the current
   TensorFlow 2.15.0 path, by running basic_pitch.inference.predict() on
   tests/resources/vocadito_10.wav against all three backends directly:

   backend	note events	note sum
   tensorflow-macos 2.15.0 (current)	28	7604.52099609375
   onnxruntime (nmp.onnx)	28	7604.51904296875
   coremltools (nmp.mlpackage)	28	7604.52099609375

3. Confirmed the fix end-to-end on real Python 3.12.0, macOS arm64 (the
   exact environment from macOS Python 3.12 dependency marker selects TensorFlow-macOS without cp312 wheels #203's repro): uv pip install --dry-run now
   resolves cleanly (installs coremltools, no tensorflow-macos attempted);
   a real install correctly reports TF_PRESENT: False, CT_PRESENT: True,
   auto-selects the .mlpackage model, and predict() produces the same 28
   note events with matching output sums as the TensorFlow 2.15.0 baseline.

4. Ran tests/test_inference.py on both the existing Python 3.11 (TF) path
   and the new Python 3.12 (CoreML) path: 8/9 pass on 3.12. The one failure
   (test_predict) is a pre-existing, unrelated issue -- librosa.get_duration(filename=...)
   was deprecated in librosa 0.10.0 and removed in 1.0.0, which is what a
   fresh librosa>=0.8.0 install now resolves to. I did not fix it here
   since path= (the replacement kwarg) doesn't exist before librosa 0.10.0,
   and basic-pitch supports librosa>=0.8.0 with no upper bound -- a
   version-safe fix is a separate, unrelated piece of work. Verified this
   failure is unrelated to this PR by reproducing it on the current Python
   3.11/TensorFlow path too (same error, same cause, present on main).

Additional Information

• Has associated issue: macOS Python 3.12 dependency marker selects TensorFlow-macOS without cp312 wheels #203