perf: avoid deep copies in comparisons and hot paths

[henryiii]

🤖 AI text below 🤖

Part of #1143

Removes hidden deep copies on hot paths. No behavior changes intended; the full test suite, mypy strict, and pre-commit all pass.

Changes

• C1 — __eq__/__ne__ lambdas in register_histogram.hpp (both the generic and multi_cell registrations), register_axis.hpp, register_storage.hpp (all three specializations), and register_accumulator.hpp used py::cast<T>(other), which casts by value and deep-copied the entire histogram/axis/storage/accumulator on every comparison. Now py::cast<const T&>(other); reference casts raise py::cast_error for foreign types exactly the same, so the False/True fallback behavior is unchanged.
• C3 — Histogram.to_numpy called the C++ to_numpy helper, which builds a full copy of the bin contents that the Python code immediately discarded (it only kept the edges). The edges are now computed in Python from the per-axis C++ edges property plus the same flow values and nextafter upper-edge nudge. The convention is replicated exactly (including category axes' 0..size edges and the regular_none/regular_uflow/regular_numpy no-nudge exceptions) and locked in by a new parametrized test that compares against the still-present C++ helper for every axis flavor, with flow=False and flow=True.
• C4 — tuple(_histograms) was rebuilt on every isinstance check in __init__, _clone, __itruediv__, and _compute_inplace_op. The set is fixed at import time (only @register reads it), so it is now hoisted into a module-level _histogram_types tuple. Also, _handle_slice now returns early for full slices [:] before calling _process_loc, whose results were unused in that case.
• C5 — serialization.to_uhi accessed the storage_type property four times (each access does a subclass-walking cast) and instantiated the class four times. It now looks the type up once and reuses a single instance (h.storage); as a side effect, the unsupported-storage error for MultiCell now reports the real nelem instead of a default-constructed instance.

Testing: pytest -n auto --benchmark-disable (1055 passed, 1 xfailed), mypy strict (no issues in 32 files), prek -a clean. A new regression test (test_to_numpy_edges_match_cpp_helper) pins the to_numpy edge convention against the C++ implementation across all axis types, including the bh.numpy regular_numpy axis.

🤖 Generated with Claude Code