true_angular is not a distance?

This may be based on my lack of understanding as covered in https://github.com/lmcinnes/pynndescent/discussions/231 but it seems like the `true_angular` function returns a similarity and not a distance, i.e.:

```python
from sklearn.datasets import load_iris

iris = np.float32(load_iris().data)

import pynndescent.distances as dist

dist.cosine(iris[0], iris[0]), dist.cosine(iris[0], iris[1])
(0.0, 0.0014208250387670596)

dist.true_angular(iris[0], iris[0]), dist.true_angular(iris[0], iris[1])
(1.0, 0.9830298038991578)
```

Edited to add that `alternative_cosine` *is* a distance, which is what is actually used for finding nearest neighbors when `metric = "true_angular"`, although it gets transformed back into a similarity via `true_angular_from_alt_cosine` at the end).

Start with the pynndescent.distances implementations of true_angular, alternative_cosine, and true_angular_from_alt_cosine, then trace how metric="true_angular" is used for nearest-neighbor search. Compare the documented and observed similarity or distance semantics, and consider the issue resolved when the public behavior is consistent and covered by a regression test.

Written by the indexing model from the issue text.