Tweak thresholds for typosquat detection

[scastlara]

These thresholds were added just as placeholders: https://github.com/elementsinteractive/twyn/blob/main/src/twyn/similarity/algorithm.py#L13

We need

• Sane defaults informed by some kind of test/experiment, selecting a false positive rate and false negative rate we are happy with.
• Exposing them through configurations.

[scastlara]

• Find optimal npm top K where the last package (at position K) is as popular as a PyPi. That's the number of packages we must download for our golden set.
• Change thesholds to not be based on two discrete sets (short words vs long words). Instead it should be a function f(x) that takes the length and returns the allowed number of typos allowed. The shape of the curve of that function idk.
• Find a maxima for the optimization problem of finding the best thresholds (and function parameters from the previous function) that returns the lower false positive rate and lower false negative rate (or just the precision recall F value is alright).
• Treat npm namespaces differently. We won't care about the package name (after /), we will consider only the namespace for typosquats IF and only if the package name is the same, e.g., asw/s3 is a typosquat, asw/e3 is not.

[sdn4z]

remaining items will be introduced in this version's minor releases (v6.x.x)