feat: add Collector storage that keeps original per-bin values

[henryiii]

🤖 AI text below 🤖

Summary

Adds bh.storage.Collector, a storage that keeps the original sample values that fall into each bin (a variable-length list per bin) instead of aggregating them. This finishes the idea started in #378 — the C++ accumulators::collector<> it builds on has since been upstreamed into Boost.Histogram and is already vendored in extern/, so this PR is the Python binding + integration.

import boost_histogram as bh
h = bh.Histogram(bh.axis.Regular(3, 0, 3), storage=bh.storage.Collector())
h.fill([0, 0, 1, 2, 2, 2], sample=[1., 2., 3., 4., 5., 6.])
h.view()   # object array: [array([1., 2.]), array([3.]), array([4., 5., 6.])]
h[0]       # [1.0, 2.0]

Each value in sample= is appended to the bin chosen by the corresponding coordinate.

Design

• Storage is dense_storage<accumulators::collector<std::vector<double>>> (double only), filled through a new fill_impl overload for the collector's accumulator_traits_holder<false, const double&> (none of the existing overloads matched).
• Because per-bin data is ragged, there is no buffer protocol. view() returns a NumPy object-dtype array (shape = axes), each element a 1-D float64 copy of that bin's values. A dedicated register_histogram specialization replaces def_buffer with an bh::indexed-based object-array builder.
• Works (all via the C++ backend, all concatenating where bins merge): integer/slice __getitem__, project, slicing/factor-rebin reduce, h1 + h2 / +=, sum, pickle, copy/deepcopy.
• Raises NotImplementedError — the object view returns copies, so anything that writes back through it is unsupported: item assignment, array/scalar arithmetic (h * 2), group rebinning, integer picking on a subset of axes, and list-based selection. Weighted and threaded fills raise too.
• Pickle stores a per-bin counts array plus a single flat values array — the offsets+content layout a future to_awkward() would need.

Out of scope / follow-ups

• Awkward Array conversion (the serialization layout is the groundwork).
• A writable ragged view (would re-enable setitem / group rebin).
• Non-double element types.

Not UHI-serializable, consistent with MultiCell.

Tests

• New tests/test_collector.py (23 tests): fill, object-array view incl. flow, scalar-arg broadcast, 2D, indexing, sum/project/reduce/factor-rebin concatenation, addition, pickle round-trip, copy/deepcopy, reset, equality, structural match, and all the unsupported-op guards.
• Full suite: 978 passed, 1 pre-existing xfail. prek -a clean (ruff, clang-format, mypy).