The build used to depend on smrt-run.jar but we don't use it anymore, so
the task "jmhTest" would fail.

We don't need to plot the accuracy for them, since we're unlikely to
replace them.

Analyzes the log and prints the distribution of accuracy

This will make the distributions we plot a little nicer.

We need that information when building JBrainy variants of programs.

It was too strict, and if it's unbalanced, we'll see it in the metrics
we compute about classification.

The script now uses 5-fold cross validation and reports the BALANCED
accuracy, which should therefore be worse if the data is too
unbalanced.

Added some code which helps with tracing the new collections:
ArrayMap, ArraySet, HashArrayList, etc.

Also centralized a big the metadata about collections (which
collection implements which interface, etc.) in
se.lth.cs.collections.meta.Collections.kt

We need that to be able to extract the features.

It's taking too long otherwise

Instead of passing arguments through gradle, we use the JMH API
directly. This is cleaner, and it's easier to find the parameters
used, I think.

We want to compare them

We don't need the others for now

I had set the wrong unit for the measurement times, now it makes more
sense if you treat it as microseconds

They called an outdated method

In some cases, 100 didn't work, but 1000 did.

The code doesn't change anything, it's just not to lose it

I used splines instead of gaussian processes, which yielded faster
training times. I also improved the evaluation steps (reporting
distribution of accuracy, and better investigations of the predictions)

We did find a benchmark with 1000 operations for which LinkedList wins
some times!

We switch to a random power of 10 for the trace sizes.

This was broken because the "measurementTime" depends of the trace size
now

Benchmark selection produces a file which stores information about
selected benchmark. THis class is to be able to reload them, benchmark
them, and produce a JMH results file which the rest of the pipeline will
understand.

Instead of using a fixed length, we use a distribution of lengths, which
still depends on the seed provided to generate the benchmark. The size
will be printed in the resulting file

We also adapt the measurementTime to be compatible with very long traces
which can happen from time to time.

Extracted function to compute options to pass to a JMH runner