Interpolate waveforms to 30kHz for classifier compatibility

[abuzarmahmood]

Fixes #193

This PR adds waveform interpolation to enable the classifier to work with data recorded at different sampling rates, as long as the pre and post spike windows are the same.

Changes

• Added interpolate_waveforms_to_30khz() function in utils/clustering.py: Interpolates waveforms from any sampling rate to match the expected 30kHz sampling rate using cubic interpolation
• Added interpolate_waveforms() method to spike_handler class in utils/blech_process_utils.py: Wrapper method that uses the spike snapshot parameters and actual sampling rate from the data
• Updated blech_process.py: Interpolates waveforms before passing them to the classifier

How it works

When data is recorded at a different sampling rate (e.g., 20kHz), the same time window (e.g., 0.5ms pre + 1.0ms post) will contain a different number of samples than at 30kHz. The interpolation function:

1. Calculates the expected number of samples at both the actual and target sampling rates
2. Creates time vectors for both sampling rates
3. Uses cubic interpolation to resample each waveform to the target number of samples
4. Returns waveforms with the same shape as if they were recorded at 30kHz

This allows the classifier, which was trained on 30kHz data, to work correctly with data from any sampling rate.

[abuzarmahmood]

Solution Overview

This implementation solves the sampling rate compatibility issue by interpolating waveforms to a standard 30kHz sampling rate before classification.

Technical Details

The solution adds a new interpolation step in the processing pipeline:

1. Waveform Extraction → Extract waveforms at native sampling rate (e.g., 20kHz, 25kHz, 30kHz)
2. Dejittering → Align waveforms by their minima
3. Interpolation (NEW) → Resample to 30kHz using cubic interpolation
4. Classification → Pass standardized waveforms to classifier

Key Implementation Points

• Cubic interpolation preserves waveform shape while resampling
• Early return optimization when actual sampling rate equals target (no unnecessary computation)
• Time-based interpolation ensures temporal accuracy regardless of sampling rate
• Minimal code changes to existing pipeline - only adds one function call before classification

Example

For a 1.5ms window (0.5ms pre + 1.0ms post):

• At 20kHz: 30 samples
• At 25kHz: 37.5 → 37 samples
• At 30kHz: 45 samples

After interpolation, all waveforms have 45 samples, matching the classifier's expectations.

Testing Recommendations

To verify this works correctly:

1. Test with data at different sampling rates (20kHz, 25kHz, 30kHz)
2. Verify classifier predictions are consistent across sampling rates for the same neural data
3. Check that waveform shapes are preserved after interpolation