report_rmses should report standard error, not standard deviation

[tbekolay]

The standard deviation is pretty meaningless right now. Standard error is more appropriate for what we're measuring.

[arvoelke]

I will also propose a bootstrapped confidence interval as a candidate!

Code from https://github.com/ctn-waterloo/best-practices/blob/f7886301c01f6a54313c5d5e5212f2c54a4e2fca/Confidence%20Intervals%20-%20bootstrap.ipynb:

def bootstrap_ci(data, func, n=3000, p=0.95):
    index = int(n*(1-p)/2)

    samples = np.random.choice(data, size=(n, len(data)))
    try:
        r = func(samples, axis=1)  # if the function supports axis
    except TypeError:
        r = [func(s) for s in samples]  # otherwise do it the slow way
    r.sort()
    return r[index], r[-index]

where func=np.mean.