Adding `from_step` option to leave the first entries

[stared]

It intends to solve #124 - as very often the first epochs have very large (and largely not meaningful) values. The idea is to use positive values so that we say start from step 5, and negative (e.g -100), so that we show up to 100 last steps. For zero we don't use the option at all.

125 has an approach for MatplotlibPlot only. I wanted to do it differently - by creating an option for MainLogger, so that it will propagate to all other elements as well (e.g. ExtremaPrinter).

• [X] add to MainLogger
• [X] make it work with MatplotlibPlot (sort of works)
• [x] make it work without pesky warnings :)
• [x] make it work with ExtremaPrinter (we want, and need, to remove value caching)
• [x] make it work in general
• [x] write tests

Right now it "sort of works":

Then later:

[stared]

Now the code

groups = {'acccuracy': ['acc', 'val_acc'], 'log-loss': ['loss', 'val_loss']}
plotlosses = PlotLosses(['ExtremaPrinter'], groups=groups, from_step=2)

for i in range(10):
    plotlosses.update({
        'acc': 1 - np.random.rand() / (i + 2.),
        'val_acc': 1 - np.random.rand() / (i + 0.5),
        'loss': 1. / (i + 2.),
        'val_loss': 1. / (i + 0.5)
    })
    plotlosses.send()
    sleep(0.3)

Runs without errors, and generates:

acccuracy
    training             (no values!)
    validation           (no values!)
log-loss
    training             (no values!)
    validation           (no values!)
acccuracy
    training             (no values!)
    validation           (no values!)
log-loss
    training             (no values!)
    validation           (no values!)
acccuracy
    training             (min:    0.778, max:    0.778, cur:    0.778)
    validation           (min:    0.973, max:    0.973, cur:    0.973)
log-loss
    training             (min:    0.250, max:    0.250, cur:    0.250)
    validation           (min:    0.400, max:    0.400, cur:    0.400)

Some auxiliary note:

• acccuracy in the example has an inaccurate spelling ;)
• groups shouldn't be in the default example
• we need to add flush mechanism even in the absence of MatplotlibPlot