fix: undefined behavior in indexed() for rank-0 histograms

[henryiii]

🤖 AI text below 🤖

Fixes #430.

For a rank-0 histogram, indexed_range::iterator::operator++ dereferenced the first slot of the per-axis index buffer, which is empty when there are no axes. That read uninitialized stack memory and, with unlucky stack contents, walked the cell iterator out of bounds — observed as a reproducible segfault on free-threaded Windows via the Python bindings (scikit-hep/boost-histogram#1153).

The fix skips the index update when the index buffer is empty; the single cell of a rank-0 histogram is still visited exactly once, so indexed(), algorithm::sum(h, coverage::inner), and algorithm::empty now all work for rank 0.

Regression tests cover indexed() (both coverages), sum, and empty on a rank-0 histogram. They pass only by luck without the fix (the uninitialized read is stack-layout dependent and needs MemorySanitizer to detect reliably), but they pin down the intended behavior.

[HDembinski]

This needs an update now.

I am surprised that a rank 0 histogram still has a single accumulator. I suppose that's consistent with numpy, but not something I remember defining as an explicit edge case.