Write a nearestpdf finder and local cache for CORE API

[khemarato]

We added (in #626 ) a title -> file name matching algo. Here we add a content matching algo and the code to use it for deduping files. Then we add the code for combining the filename signal with the content signals to get an overall matching algo capable of saying whether we already have a given CORE API work.

This PR also adds the local SQLite manager which will oversee pulling data from the CORE API.

[netlify]

✅ Deploy Preview for obu ready!

Name	Link
🔨 Latest commit	1ba79ff
🔍 Latest deploy log	https://app.netlify.com/projects/obu/deploys/698daa883b50cd00083f4685
😎 Deploy Preview	https://deploy-preview-629--obu.netlify.app
📱 Preview on mobile	Toggle QR Code... Use your smartphone camera to open QR code link.

---

To edit notification comments on pull requests, go to your Netlify project configuration.

[khemarato]

After loading most of the PDFs and finding their nearest neighbors, here is a plot of the cosine similarities zooming in on the difficult range:

Here is the entire histogram, trying to fit a bimodal binomial distribution:

This suggests a threshold value of 0.965.

[khemarato]

For the final combination of the title match scores and the document cosine similarities:

Fitting a Logistic Regression to the above scatterplot yields the equation:

$$z = -6.04 + (5.86 \times \text{ContentSim}) + (3.90 \times \text{TitleSim})$$ $$P(\text{match}) = \frac{1}{1 + e^{-z}}$$

Adding in a hundred manual examples from PDF similarities as well, that version was more conservative, giving a cutoff equation of 6.33 ContentSim + 3.05 TitleSim > 6.73 when doing the regression: trusting the Content sim a little less and the title sim significantly less.

After playing around on Desmos I found a good line between the two regressions at 18 x + 10 y - 19 > 0. This equation has the nice property that it requires a ContentSim > 0.5 and Title P > 0.1 which makes pruning straightforward.

Playing around with normalizing the Logistic curve, I found that dividing the Z score by 3 gives reasonable P values in the range [0.5, 0.953) over the output domain.