Allow tolerance for `unique`

[dpsanders]

When working with Vectors of floating-point objects, it would be useful to be able to specify a tolerance for unique.

[StefanKarpinski]

Allowing a tolerance naively requires an O(n^2) algorithm instead of the O(n) hash-based algorithm that unique uses. There may be clever ways to make it more efficient (e.g. sorting and bucketing) but I'm not sure that level of cleverness belongs in Base. There's also the issue that adding a tolerance causes the equivalence relation to cease to be an equivalence relation anymore.

[mbauman]

For some purposes along these lines you might just want hist. Er, fit(Histogram, _).

[dpsanders]

The use case I had in mind is finding fixed points roots of functions, e.g. using the Newton method.
Starting from different initial conditions leads to possibly repeated or possibly different fixed points.

But I see that probably this does not belong in Base.

[yuyichao]

Dup of #19147 btw