detect_silence finds separate slices of silence and in a last step
combines subsequent silent slices into ranges of continuous silence.
The added tests specifically ensure the correct function of this
combination step.

Previously, detect_silence would collect all slices of min_silence_len
in a list, then processed that list to merge subsequent slices
into continuous silent ranges.
This change performs the merging immediately when silence is detected
for a slice, eliminating the need for a second pass over and the memory
overhead associated with the internal list of silent slices.

Using numpy to compute RMS for silence detection to reduce
redundant computation (and benefit from numpys highly optimized
implementation) compared to previous implementation of
detect_silence.

Some caveats:
- adds numpy as new dependency
- previously RMS values where rounded down to the next integer;
  this is now not the case anymore, resulting in borders of silence
  ranges to possibly vary slightly compared to previous implementation