Maybe sorting `BigFloat`s could be faster?

[LilithHafner]

Sorting benchmarks vs. 1.8 are mostly wins or runtime wins with increased allocations.

Of the four exceptions, three include BigFloat. We/I should look into that before the 1.9 release.

[LilithHafner]

This script might give some interesting results:

using BenchmarkTools
v = [BigFloat(rand()) for _ in 1:10_000];
@btime sort($v);

mutable struct Pu238
    location::Tuple
    measurements::Vector{Integer}
end
Base.isless(a::Pu238, b::Pu238) = hash(a.location) + hash(a.measurements) < hash(b.location) + hash(b.measurements)
unstable = [Pu238((5,3,2), rand(Int, rand(1:10))) for _ in 1:10000];
@btime sort($unstable);

struct Pu244
  location::NTuple{3, Int}
  measurements::Vector{Int}
end
Base.isless(a::Pu244, b::Pu244) = hash(a.location) + hash(a.measurements) < hash(b.location) + hash(b.measurements)
stable = [Pu244((5,3,2), rand(Int, rand(1:10))) for _ in 1:10000];
@btime sort($stable);

[KristofferC]

I don't think a 5% runtime regression + a non drastic memory increase for sorting bigfloats is enough for a milestone. Would of course be good to fix it though.