Honor the dim argument when histogramming by a dense coord - #3959

Fixes #3958.

ADR 0018 defines dim as the dimensions to be replaced, so dim=() must preserve all input dims and add a new one. bin, group and nanhist do this; hist did not, whenever the coord used for histogramming is a dense (outer) coord of binned data rather than an event coord. The tabulated ADR row (x,) | (x,) | (x,y) | .hist(y=y, dim=()) was violated, and the same call with a multi-dimensional coord raised DimensionError instead — including the exact example the hist docstring uses to advertise dim=().

In that situation make_histogrammed summed the bins and handed the result to the C++ dense histogram, which always replaces all dims of the coord by the new dim. erase only drove a flatten beforehand, so dims outside it could not survive. Bin instead when erase does not already cover the coord's dims, which is what nanhist (bin followed by nansum) has always done. The dense fast path is kept when erase does cover them, so the common dim=None case is unaffected — hist(t=10) over 5M events stays at 2 ms rather than moving event data.

This is not a regression from a recent optimization: the fallback dates from #3020 (2023) and predates the ADR. It went unnoticed because test_op_on_binned_with_explicit_dim_arg_yields_expected_output_dims leaves the event coord in place, so the fallback branch was never exercised. This PR adds the same table without the event coord.

Two consequences beyond the bug itself:

• One of the xfail cases from Fix OOM in multi-dim hist depending on argument order #3949 now passes and is strengthened to check the values against a reference that broadcasts the outer coords onto the events, as requested in review. The remaining xfail there (re-binning a dim drops coords defined over it) is unaffected, so hist depends on argument order when an edge coord is defined on a re-binned dim #3952 stays open.
• The reordering guard added in Fix OOM in multi-dim hist depending on argument order #3949 is dropped. It existed because histogramming by an outer coord erased that coord's dims, so which edges came last changed which dims the result had. That is no longer true.

Not in scope: dense input

Dense (non-binned) data has the same gap on a separate code path, which this PR does not touch. There bin, group and nanhist are non-conforming too, so routing through bin is not an option and a different mechanism is needed. Tracked in #3960, reproduced here so the boundary of this PR is clear.

bin, group and nanhist agree in every cell, so they share a column. ! marks a result that violates ADR 0018. Kind is the dividing line: degenerate means the coord is constant across everything being summed (coord.dims ∩ dim = ∅), so each preserved position feeds exactly one z bin and the output is the input scattered one-hot. Meaningful means the coord varies over a summed dim, so the histogram does real work.

Binned input conforms in all twelve equivalent cells after this PR, degenerate ones included. The asymmetry is intended: binned data composes the degenerate case with useful edges, which is exactly what #3958 was, whereas dense data cannot.

Stacked on #3956, which is stacked on #3949. Review/merge those first.

Test plan

• New parametrized test covering the full dim table for bin, group, hist and nanhist on binned data whose histogramming coord exists only as an outer coord, over 1-D and 2-D coords. All three parametrizations fail on the base branch.
• The hist docstring gains a dim=() example, so the documented behaviour is now checked by the doctests.