LLVM-Error when using mahalanobis metric with larger datasets

Hello!

First: Thanks for this splendid project! It's looking great, especially the semi-supervised part.

When I try to use Mahalanobis-distances with larger datasets, so the < 4000 codepath is not taken, I get an LLVM IR parsing failure. The other codepath works just fine.

Steps to reproduce:

Version: umap-learn-0.3.2

import numpy as np
from umap import UMAP

matrix = np.random.rand(5000,50)
umap = UMAP(n_components=2, n_neighbors=30, metric='mahalanobis', metric_kwds={'V': np.eye(50)})

umap_model = umap.fit_transform(matrix)

Resulting in

Failed at nopython (nopython mode backend)
LLVM IR parsing error
<string>:1121:137: error: invalid use of function-local name
  %".786" = extractvalue [1 x {i8*, i8*, i64, i64, double*, [2 x i64], [2 x i64]}] [{i8*, i8*, i64, i64, double*, [2 x i64], [2 x i64]} %".785"], 0

Full stracktrace (redacted for brevity) here:
https://gist.github.com/johanbev/1918108e65f014600b2e44affcc35fee

Start by running the reported umap-learn 0.3.2 reproduction with a 5000x50 NumPy matrix and Mahalanobis metric, then compare it with the smaller-dataset path. Use the linked full stacktrace to locate the failing nopython/LLVM path; done means the larger-dataset case completes without an LLVM IR parsing failure.