Raise a descriptive error when unscaled integer audio is passed (fixes #407)

[deekshaNVIDIA]

Summary

Fixes #407.

In several reported cases (#406, #390, #369), users passed integer PCM data (e.g. int16) that had been converted to float32/float64 without rescaling it into the [-1.0, 1.0] range that Pedalboard expects. Scale-invariant plugins (Reverb, Delay) appear to work, while others (Compressor, Limiter, PitchShift) fail in confusing ways.

This adds a small check in the shared C++ process() entry point in pedalboard/process.h — the single code path used by pedalboard.process(...), Plugin.process(...)/Plugin.__call__, and Pedalboard.__call__ (via Chain). If every sample is an exact integer and the peak magnitude exceeds 1.0, it raises a ValueError explaining the likely cause and the fix.

Behavior

• Raises a descriptive ValueError for buffers that look like unscaled integer PCM data (e.g. (audio * 32767).astype(np.int16).astype(np.float32)).
• Does not affect legitimate float audio. The check returns on the first fractional or non-finite sample, so for real-world audio it adds only a handful of instructions and never allocates.
• In-range integer-valued buffers (digital silence, a full-scale ±1.0 square wave, DC at 1.0) are still accepted, since those are valid audio.

Why this location

process() in process.h is the single chokepoint that all Python-facing processing entry points route through, so the validation covers every case (Pedalboard, individual plugins, and the module-level process function) with one implementation. The check runs while the GIL is still held, before the heavy processing begins.

Notes on testing

I validated the detection heuristic numerically against the released pedalboard wheel across many signal types (sine waves, random noise, silence, full-scale DC/square waves, hot non-integer signals, empty and stereo buffers): zero false positives, and the documented int16-as-float32 case is caught (its peak of ~29490 otherwise reaches the Limiter).

tests/test_integer_input_validation.py is added to cover:

• int16-as-float32 and -as-float64 raising ValueError via Pedalboard, a single Plugin, and pedalboard.process.
• Correctly-rescaled integer data being accepted.
• A range of legitimate float signals not being flagged.

I was unable to compile the native extension locally (no C++ toolchain on my machine), so I have not run the compiled test suite end-to-end — CI will build and run the new tests. ruff, black (line length 100), and flake8 all pass on the new test file.