Faster Hjerson bootstrap resampling

[martinpopel]

Currently, bootstrap resampling is done schematically

for( $i = 0; $i < 1000; $i++ ) {
    $metric->init();
    for( $j = 0; $j < $count; $j++ ) {
    $rand = mt_rand( 0, $count - 1 );
    $metric->addSentence( $sentences[ $rand ]...);
    }
   $samples[] = $metric->getScore();
}

That is we calling $metric->getScore() 1000 times. So for Hjerson we run the external Python script 1000 times. For BLEU, we precompute sentence-level (and n-gram level) statistics in advance, so get_score() and add_sentence() is much faster. We would need to get sentece-level scores from Hjerson.
We should also compute 5 Hjerson metrics in one run (instead of five runs).

[martinpopel]

Partially solved in #50