Add sliding window operations to speasy.signal

[jeandet]

Description

Add a sliding window mechanism to speasy.signal that applies a function over a moving time window on a SpeasyVariable. This enables common time-series analysis operations like moving averages, sliding FFTs (spectrograms), running integrals, and arbitrary user-defined windowed computations.

Use cases

• Moving average — smooth noisy time series over a configurable window
• Sliding FFT / spectrogram — compute power spectral density in a moving window, producing a time-frequency representation
• Running integration — fluence, accumulated flux over a sliding interval
• Custom operations — any user-provided function applied per window (e.g. variance, percentiles)

Sketch of possible API

from speasy.signal.windowing import sliding_window

# Moving average
smoothed = sliding_window(var, window_size=np.timedelta64(30, 's'), func=np.mean)

# Sliding FFT (returns a different shape — time x frequency)
spectrogram = sliding_window(var, window_size=np.timedelta64(10, 's'), func='fft', step=np.timedelta64(5, 's'))

# Custom callable
result = sliding_window(var, window_size=np.timedelta64(1, 'm'), func=my_function)

Design considerations

• window_size as np.timedelta64 or float (seconds)
• Optional step parameter for non-overlapping or partially overlapping windows (defaults to one sample = fully overlapping)
• Built-in fast paths for mean (cumsum trick, O(n)) and fft (stride tricks + vectorized FFT)
• Arbitrary callable fallback for custom functions
• Output shape depends on the function: same shape for mean/integrate, time x frequency for FFT
• Should handle lists of variables like the rest of speasy.signal

[brenard-irap]

I would suggest adding two additional optional parameters:

• window_alignment: "center", "left", or "right" to specify whether the window is centered, left‑aligned, or right‑aligned relative to the reference time.

• ref_time: Explicit definition of the list of reference times that serve as the anchor points at which each sliding window is evaluated.