57.3 -> (180.0 / M_PI)

This gets rid of all the M_PI, 60 and 3600 factors everywhere.

sqrt(2 * log(2) / M_PI) = 0.6642824703

See IEEE Standard 952-2020 for Specifying and Testing Single-Axis
Interferometric Fiber Optic Gyros
(https://doi.org/10.1109/IEEESTD.2021.9353434), Table B.2 (p. 68)

This fixes the problem that often the optimization would abort with
NO_CONVERGENCE after the default 50 iterations, which could be seen with
summary.FullReport(). Now it converges every time.

... so that it follows the standard Q N B K R order.

sqrt(calcSigma2(C_Q_, C_N_, C_B_, C_K_, C_R_, 1 ) and getN() produce
almost the same result (because N dominates the function at tau=1), but
using the estimated N is correct.

Sources:

- "By multiplying the noise density (ND) by the square root of the
  sampling rate (SR), the noise standard deviation at that rate can be
  recovered."
  (https://www.vectornav.com/resources/inertial-navigation-primer/specifications--and--error-budgets/specs-imuspecs)
  sqrt(SR) = sqrt(1/freq), so we have to divide by sqrt(freq).
- Also see the function generate_signal here: https://github.com/nmayorov/allan-variance/blob/master/example.ipynb

Modified generate_signal:

```python
def generate_signal(n, dt, q_white, q_walk, q_ramp, random_state=0):
    rng = np.random.RandomState(random_state)
    white = q_white / dt**0.5 * rng.randn(n)
    walk = q_walk * dt**0.5 * np.cumsum(rng.randn(n))
    ramp = q_ramp * dt * np.arange(n)
    return white + walk + ramp
```

The discrete-time standard deviations are only used for generating
simulated data from the continuous-time standard deviations.