Bug: Flaky behavior around entrypoint deletion in HNSW index

[etiennedi]

HTR

Create simple zero-shot (#1603) classification with two sources and two targets. Run 100 times.

EB

All 100 runs succeed

AB

Typically before the 100 runs are over, we can see a nil pointer deference panic occur because the *vertex{} object of a newly changed entrypoint is nil

Initial Idea

It seems there is a missing lock somwhere when retrieving the EP.