Improve "time of flight" fit

The time of flight fit currently consist in the following algorithm:

a low pass filter (moving average) to remove noise and the carrier
a discrete derivative
an argmax, to find the position of envelope rise

https://github.com/qiboteam/qibocal/blob/b0c0ef7de08ef35d3f5b8a0140349d3c95724b58/src/qibocal/protocols/signal_experiments/time_of_flight_readout.py#L96-L102

However, this is definitely committed to a given pulse shape, and prone to signal distortions.

A more stable alternative would be to inspect the macroscopic properties of the signal, of which we need only one: the first moment, i.e. its (horizontal) average (expected value, as in distributions).

In order to extract it, the baseline value should be subtracted (otherwise the average would be affected by the centering of the acquisition window), and in principle the carrier should be removed anyhow, to avoid oscillating about 0. This second step could be easily replaced by an absolute value, since the (horizontal) average will be asymptotically insensitive to high-frequency oscillations. In the end, the proposal will be something like:

baseline = signal[:init].mean()
shift = (signal - baseline).abs()
mean = (shift / shift.sum() * np.arange(len(shift))).mean()

to be compared to the mean of the signal sent on readout.

However, to work as intended, this requires the following assumptions:

no hysteretic behavior on the channel (the baseline should be the same before and after the signal)
the full signal should be acquired

While 1. is a sufficiently general and reasonable assumption, 2. is currently plainly false while using Qibolab. Indeed, the acquisition is done in sync with the readout pulse, cutting the end of the signal. This could be solved by defining the acquisition window independently of the readout pulse.

qiboteam / qibocal

Improve "time of flight" fit #876