Avoid quadratic Quantiles behaviour on datasets with many duplicates.

[ionik0]

Fixes issue no #3789.

Sending this because of your comment on the issue about maybe having a look if the fix were easy enough to understand. Happy to close it if you'd rather not take this on.

Problem

partition places values equal to the pivot on the low side of the partition point. On a dataset where one value is very common, the range being searched shrinks by only one element per iteration, so selectInPlace becomes quadratic.
The median of 200,000 equal values takes about 5.9 seconds.

Why not simply replace the partition

I tried that first, three ways: a Dutch-national-flag partition, Bentley-McIlroy ,and a Hoare partition that splits equal values across both sides. All three fix the duplicate case, and all three cost 20-25% on duplicate-free input, measured over 400 independent datasets per configuration. The existing countdown loop optimizes better than any partition with data-dependent loop bounds, so the cost looks inherent rather than an artefact of my implementations.

So this keeps the existing partition for the common case and gives it a budget of iterations comfortably larger than well-behaved data needs. Only once that budget is exhausted do we switch to a three-way partition, which gathers every value equal to the pivot into a single run so the run can be excluded in one step.
Duplicate-free input never reaches the fallback and pays only one decrement and one comparison per partition step.

Numbers

Median of 200,000 values, end to end through the public API:

dataset : before --> after

all values equal : 5875.76 ms --> 2.65 ms
5 distinct values : 208.67 ms --> 1.57 ms
100 distinct values : 2.37 ms --> 1.51 ms
gaussian (all distinct) : 1.52 ms --> 1.61 ms

Tests

Every existing Quantiles test dataset is Gaussian, so duplicates never arise and this case was uncovered. The new tests are sized so they would take minutes rather than seconds against the old code.

Notes

NaN cannot reach the selection code, since containsNaN short-circuits first,
so the three-way partition does not need to handle it.
The budget constant (2 * log2(size) + 4) is the one tuned number here; happy
to justify or rename it.
I have not touched android/, per CONTRIBUTING.