This commit speeds up my_partialsort! to be faster than it was before, but partialsort! is slower than any version of my_partialsort!

julia> using Random; Random.seed!(1); x = rand(1000); hash(x)
0x13843e5a90ce1e51

julia> @Btime partialsort!(copy($x), 500);
  3.838 μs (3 allocations: 7.88 KiB)

julia> @Btime my_partialsort!(copy($x), 500);
  1.946 μs (3 allocations: 7.88 KiB)

…ble IEEE opt. because it makes the kernel slower)

julia> using Random; Random.seed!(1); x = rand(1000); hash(x)
0x13843e5a90ce1e51

julia> @Btime partialsort!(copy($x), 500);
  2.671 μs (3 allocations: 7.88 KiB)

julia> @Btime my_partialsort!(copy($x), 500);
  2.000 μs (3 allocations: 7.88 KiB)

julia> @Btime partialsort!(copy($x), 500, lt=<);
  1.733 μs (3 allocations: 7.88 KiB)

…mize scalar arithmetic to avoid >100% regression when taking fallback

…mance bugfix)

…tes, and remove type instability in crit-fail path

…rrible)

… short-circuit to avoid dispatch overhead

…sized input success rates)

… drop max trials to 3 and adjust comments

… readability

… tests

There are some pretty compelling reasons to use PartialQuickSort instead of ScratchQuickSort as a base case for BracketedSort.
However, PartialQuickSort can segfault for invalid but seemingly innocuous lt functions.
I don't want to talk about introducing segfaults to the default sorting algorithms or introducing a regression to PartialQuickSort in this PR
Revert this commit it you want `partialsort!(::Vector{Int}, ::Int)` not to allocate.