fix: pin note_seq==0.0.3, drop tensorflow-datasets (training only) to escape protobuf triangle

fix: use protobuf==4.26.1 to satisfy both TF builder and pb2 runtime_version needs

fix: patch ddsp cloud import to avoid protobuf conflict, drop google-cloud-storage

fix: add google-cloud-storage and apache-beam (remaining ddsp deps)

fix: quote version specifiers in RUN pip install to fix shell syntax error

fix: add all ddsp deps from setup.py to avoid whack-a-mole missing modules

fix: add gin-config (missed ddsp dep)

fix: force-reinstall numpy==1.23.5 at end to fix TF bfloat16 error

fix: force-reinstall protobuf==3.20.3 at end to prevent upgrade by note-seq/gradio

fix: pin protobuf==3.20.3 for tensorflow-cpu==2.12.0 compatibility

fix: add tensorflow-cpu==2.12.0 (missed due to ddsp --no-deps)

fix: install midi-ddsp --no-deps with cleared constraint to bypass numba conflict

fix: install each dep separately to avoid pip resolution-too-deep

fix: pre-install crepe with --no-build-isolation for pkg_resources

fix: use pip constraints file to pin numba/llvmlite globally

fix: install crepe with --no-build-isolation to get pkg_resources

fix: numba/ddsp compat via BeatNet pattern - numba 0.57 pre-built wheel

use python 3.10 for gradio 5 compatibility

initial MIDI-DDSP HARP deployment