Make tensorflow an optional dependency (vus[models] extra)

[martinuray]

This PR is addressing issue #6 .

TensorFlow was a mandatory dependency for the whole package, even though it's only needed by three of the baseline detectors under vus.models: lstm.py, AE.py, and cnn.py. The metrics themselves, which are the actual contribution of the paper, have no TensorFlow or GPU dependency at all.

This meant every pip install vus pulled in TensorFlow, which is a common source of dependency conflicts (TensorFlow/PyTorch/CUDA version mismatches) in projects that don't otherwise use it, even when the only thing needed from vus is the evaluation metrics.

This PR moves tensorflow out of install_requires and into an optional models extra. pip install vus now installs a lightweight package with just the metrics, and pip install vus[models] additionally installs TensorFlow for the three baseline detectors that need it.

This is technically a breaking change for anyone relying on a plain pip install vus to also pull in TensorFlow implicitly, but should be a net improvement for most users, who only need the metrics and now get a lighter install with one less source of dependency conflicts.