get_unit_adaptive_window leads to recursion errors if there are no peaks.

[zm711]

In the adaptive window code it tries to find the win_size by taking the initial threshold set by the user and if that doesn't work then it divides by 2 to try a smaller threshold and continues. But if there is no threshold then we end up with reaching the maximum recursion depth because 0/2 will keep being 0. I don't know this bit of the code well enough to know what our base condition should be in the case that there are no peaks (maybe if a unit with bad correlograms?) so I'm no completely sure how to fix this. What do think @yger since you've updated this extensively.

function below for convenient check:

spikeinterface/src/spikeinterface/curation/auto_merge.py

Lines 717 to 755 in 6fde997

	def get_unit_adaptive_window(auto_corr: np.ndarray, threshold: float) -> int:
	"""
	Computes an adaptive window to correlogram (basically corresponds to the first peak).
	Based on a minimum threshold and minimum of second derivative.
	If no peak is found over threshold, recomputes with threshold/2.
	Parameters
	----------
	auto_corr : np.ndarray
	Correlogram used for adaptive window.
	threshold : float
	Minimum threshold of correlogram (all peaks under this threshold are discarded).
	Returns
	-------
	unit_window : int
	Index at which the adaptive window has been calculated.
	"""
	import scipy.signal
	if np.sum(np.abs(auto_corr)) == 0:
	return 20.0
	derivative_2 = -np.gradient(np.gradient(auto_corr))
	peaks = scipy.signal.find_peaks(derivative_2)[0]
	keep = auto_corr[peaks] >= threshold
	peaks = peaks[keep]
	keep = peaks < (auto_corr.shape[0] // 2)
	peaks = peaks[keep]
	if peaks.size == 0:
	# If none of the peaks crossed the threshold, redo with threshold/2.
	return get_unit_adaptive_window(auto_corr, threshold / 2)
	# keep the last peak (nearest to center)
	win_size = auto_corr.shape[0] // 2 - peaks[-1]
	return win_size