Checklist

• The title and commit message(s) are descriptive.
• Small commits made to fix your PR have been squashed to avoid history pollution.
• Tests have been added for new features or bug fixes.
• API of new functions and classes are documented. (N/A — bug fix only, no API change)

Description

xt::roll enters an infinite loop (and then divides by zero) when the input
has a zero-length axis:

• No-axis overload roll(e, shift) when e.size() == 0:
  while (shift < 0) shift += flat_size; loops forever with flat_size == 0,
  then shift %= 0 is undefined behaviour.
• Axis-aware overload roll(e, shift, axis) when the rolled axis has length 0:
  same issue on axis_dim == 0.

Fix

For a shape $s = (s_0, \dots, s_{n-1})$,
if $\exists k: s_k = 0$, then $\mathrm{roll}(a, \mathrm{shift}, \mathrm{axis}) = a$,
i.e. roll is the identity.

In both overloads, after constructing cpy = empty_like(e), return cpy
early when cpy.size() == 0. The no-axis overload's hand-rolled
std::accumulate(shape, 1L, multiplies<size_t>()) is replaced by cpy.size()
(numerically equivalent, and O(1) on contiguous layouts).

When the rolled axis is non-empty but another axis has length 0
(e.g. shape={3,0,4} rolled along axis 0), the original code didn't loop —
detail::roll's inner loops became no-ops because the stride product
collapses to 0 — but the recursion was wasted work. The widened guard
cpy.size() == 0 skips it cleanly. Observable output is unchanged.

Tests

Three regression cases in test_xmanipulation.cpp, one per code path:

All test_xmanipulation tests pass: 28 cases, 188 assertions. Verified
against numpy.roll, which returns a same-shape empty array for these inputs.