KeyError when using metric="precomputed"

Nobody has claimed this yet.

I am trying to use a precomputed square distance matrix when using the reducer and I am running into a strange KeyError: 'precomputed'

When I do the following:

```
reducer = umap.UMAP(n_neighbors=15, local_connectivity=1, n_components=2, metric="precomputed", random_state=0)
distance_matrix = np.random.rand(SIZE, SIZE)
output = reducer.fit_transform(distance_matrix)
```

Everything works fine.

But when I use Pytorch to create the matrix like so:

```
reducer = umap.UMAP(n_neighbors=15, local_connectivity=1, n_components=2, metric="precomputed", random_state=0)
distance_matrix = torch.randn(SIZE, SIZE).numpy()
output = reducer.fit_transform(distance_matrix)
```

I get a KeyError: 'precomputed'

What could be the potential cause of this? The memory layout of Pytorch tensors is exactly the same as that of numpy, and they are convertible to each other using the same allocated memory.

Start with the reported reducer.fit_transform reproducer and compare the NumPy-generated and torch.randn(...).numpy() inputs while using metric="precomputed". Trace the precomputed-metric lookup that raises KeyError; done means both input paths behave consistently without that error.

Written by the indexing model from the issue text.