Allow custom features

[carlthome]

Trying to implement a new feature by implementing Features but unable to use it with msaf.run.process because there's a hardcoded list of feature names in _preprocess in the SegmenterInterface that's not exposed to the user (as far as I can see).

Maybe I'm missing something in the docs but think this check could simply be removed, and the try/except covers any potential problems on its own.

Bit unsure of this though. Gladly take help!

[carlthome]

Additionally, I was unable to create a subclass of msaf.features.Features (trying EnCodec embeddings) to register in collection mode due to how joblib results in cleared out msaf.features_registry dicts per worker (bit of a rabbit hole this: https://joblib.readthedocs.io/en/latest/parallel.html#serialization-processes).

Added a fix on this PR.

[urinieto]

Looks good to me. Could you make sure that the documentation is up to date for this to work? Here's the part of the documentation we should make sure it still works: https://pythonhosted.org/msaf/features.html#adding-new-features-to-msaf

After this is checked, I'll merge the PR.

[urinieto]

Seems like tests are not passing. Will merge after this is fixed. Thanks @carlthome

[carlthome]

Fix for CI failures in https://github.com/urinieto/msaf/pull/143

[carlthome]

Will merge after this is fixed.

Would also like to demonstrate that this actually works first, perhaps by adding a simple example to examples/

[urinieto]

Thanks! Seems like tests are still not passing, will wait before merging.

[carlthome]

Ready to go in now. Manually tested to work in both single file mode and collection mode.

[carlthome]

Although maybe https://github.com/urinieto/msaf/pull/143 should go in first to get down the size of this diff.