Optimize matrix_to_mid performance

[floriangardin]

Musiclang score to midi takes too long (~0.2 ms / note on a macbook).

Here is how the time decompose in musiclang.write.out.midi_utils.matrix_to_mid function on a ~6000 notes score :
df creation 0.008s
df copy 0.0004s
merge_continuation_to_previous_note 0.20s
prepare_df_for_events 0.15s
apply_events 0.69s

Would be a good improvement to find a way to accelerate the three last functions.