Fix MIDI overload resolution for ArrayLike support (#450)

[gijzelaerr]

Summary

• Fixes the macOS/Windows test failures introduced by Allow passing any ArrayLike object instead of just numpy.ndarray. #450 (psobot's ArrayLike support PR)
• The root cause: Allow passing any ArrayLike object instead of just numpy.ndarray. #450 changed ExternalPlugin binding lambdas from const py::array to py::object, which broke pybind11's overload resolution between the audio process/__call__ and MIDI renderMIDIMessages overloads
• When instrument plugins were called with MIDI messages (e.g. plugin(notes, 5.0, sr, reset=False)), the py::object audio overload matched first, causing ensureArrayLike() to fail on non-array MIDI message lists with: TypeError: Pedalboard only supports 32-bit and 64-bit floating point audio for processing.
• The fix restores const py::array as the parameter type in the ExternalPlugin binding lambdas (AbstractExternalPlugin, VST3Plugin, AudioUnitPlugin) so pybind11 correctly rejects non-array inputs and falls through to the MIDI overload
• The base Plugin class and standalone process() function retain py::object since they have no MIDI overloads to conflict with
• pybind11 still auto-converts array-like objects (torch tensors, JAX arrays, etc.) to py::array via the buffer protocol and __array__ interface during argument dispatch

Based on @psobot's work in #450.

Test plan

• Linux tests should continue to pass (no instrument plugins, so no MIDI dispatch)
• macOS tests with instrument plugins (Magical8BitPlug2) should now pass — MIDI messages correctly route to renderMIDIMessages instead of the audio process path
• Windows tests with instrument plugins should similarly pass
• ArrayLike support (torch tensors, etc.) still works for audio processing via the base Plugin class and standalone process() function

🤖 Generated with Claude Code