Fix signed version of radix_sort

[schnellerhase]

Fixes bug in radix_sort, reported by @jorgensd, for signed integers where bitwise comparison failed and resulted in either infinite loops or unsorted result.

Fixes it by radix sort only working on unsigned ints internally and inputs with signed getting shifted to corresponding unsigned int ranges.

[IgorBaratta]

Why not define the projection outside and only apply it when there's signed integers?
Is there any performance overhead introduced by always performing this conversion?

otherwise, LGTM.

[schnellerhase]

Good catch - allows for making the logic quite a bit more expressive, thanks.

[schnellerhase]

Reusing schnellerhase#24.

On main:

Benchmark                               Time             CPU   Iterations
-------------------------------------------------------------------------
BM_create_box/10/iterations:5        4.08 ms         4.04 ms            5
BM_create_box/50/iterations:5         530 ms          519 ms            5
BM_create_box/100/iterations:5       4640 ms         4627 ms            5

On this branch:

-------------------------------------------------------------------------
Benchmark                               Time             CPU   Iterations
-------------------------------------------------------------------------
BM_create_box/10/iterations:5        4.72 ms         4.71 ms            5
BM_create_box/50/iterations:5         551 ms          551 ms            5
BM_create_box/100/iterations:5       5038 ms         5033 ms            5

Has some impact.

[schnellerhase]

Best guess would be that we now always max out the iteration count max_value since we set the most significant bit. Any ideas how to avoid that?

[schnellerhase]

We only need to do the sign flip in case we detect a negative number, maybe check at runtime and use the unmodified projection if no casting is necessary?

[schnellerhase]

After optimisation for first bit (ignore for order if all set)

-------------------------------------------------------------------------
Benchmark                               Time             CPU   Iterations
-------------------------------------------------------------------------
BM_create_box/10/iterations:5        3.94 ms         3.90 ms            5
BM_create_box/50/iterations:5         479 ms          479 ms            5
BM_create_box/100/iterations:5       4496 ms         4495 ms            5

[jorgensd]

@IgorBaratta is this ready to me merged?

[IgorBaratta]

@IgorBaratta is this ready to me merged?

LGTM!