evaluate `filter_complex`

[faroit]

currently stempeg loads stems by

• extracting stream-by-stream and convert to wav
• load wav files and merge to tensor

For each stream extraction a os.call is used. For small files this could slows down stempeg loading significantly.

Alternatively we can also use ffmpegs amerge filter to merge several input stream into a single stream with multiple channels.

eg.

-filter_complex', "[0:0][0:1][0:2][0:3][0:4] amerge=inputs=5 merges the 5 streams in a stem file into a 10channel stream. This reduces os.calls and should make reading faster... given the filter implementation is efficiently done in C.

[faroit]

It turned out that complex filters are slow.