feat(lyra): compute levenshtein metric within a given tolerance

[jkomyno]

As discussed privately, I have noticed that the levenshtein distance is currently only used as a comparison metric to check whether two strings have edit distance less than a given tolerance threshold.

Computing the levenshtein metric in the general case takes time O(max(m, n)) and space O(max(m, n)), where m and n is the length of the two input strings. When we're only interested in retrieving the exact edit distance when it's below or equal to a given tolerance threshold, more time-efficient methods exist.

Given

let term: string;
let word: string;
let tolerance: number; // small, non-negative integer

this PR turns the comparison

if (levenshtein(term, word) <= tolerance) {
  // ...
}

into the following:

if (boundedLevenshtein(term, word, tolerance).isBounded) {
  // ...
}

The core algorithm this PR takes inspiration from has already been developed in the talisman project, which is MIT licensed.

It'd be interested to see how the changes introduced by this PR will influence the benchmarks. I expect a marginal performance increase as the size of the inputs grows.

Alternative not considered in this PR

As the size of the input scales, computing an exact comparison metric may not be needed. Luckily, the literature offers a plethora of polylogarithmic approximation algorithms to choose from. Should you be interested in this slightly less precise alternative, I'd suggest to take a look at Andoni et al.'s method.

[micheleriva]

Hi @jkomyno,
Thank you so much for this PR, this is brilliant!

About the benchmarks, I forgot we removed the benchmark script from this repo, so I'll run them manually and update you on the results