Optimize numeric vectors in Cereal, YAS, and Boost binary

[Altren]

Cereal, YAS, and Boost.Serialization binary currently serialize numeric vectors one element at a time. Transfer their contiguous numeric payloads in one archive operation while preserving the adapters’ existing array lengths and parent metadata.

• Cereal retains the size_t length prefix and uses typed binary_data, including per-element byte swapping for portable archives.
• YAS retains the size_t prefix and uses block I/O for its existing uncompressed host-endian binary archive.
• Boost specializes only binary_iarchive / binary_oarchive, retaining the uint64_t length prefix and using load_binary / save_binary. Other archive types keep the generic parser.
• vector<bool>, Vectorstring, and non-arithmetic element types keep their existing paths. Schema generation remains inherited from VectorParser.

Validation:

• Four additional numeric-vector tests pass with CEREAL_SIZE_TYPE=uint32_t.
• New shared tests compare bytes against the original element-by-element parser and read in both directions. They cover integer widths/signs, float/double, empty vectors, boundaries, negative zero, infinities/NaNs, nested arrays, maps, optional/variant values, NoFieldNames, truncated payloads, and excessive lengths.
• Cereal tests exercise both little- and big-endian output. Boost tests verify the text-archive fallback.
• All three large numeric-vector fixtures have identical before/after bytes and successfully read baseline output. All 60 corresponding benchmark case runs complete successfully.

Added bulk_numeric_vectors to the existing all-format benchmark executable and Linux/macOS/Windows benchmark summaries. The results below cover a deterministic std::vector<uint64_t> with 1.5 million elements, using ordinary reflect-cpp APIs without custom serializers or field annotations.

To isolate the incremental benefit, both measured variants include #738 and the reserve-capacity changes from #739 (ce5db49). The optimized variant adds only this PR’s bulk-I/O changes. This branch itself is based on upstream main and does not depend on #739.

Format	Dataset	Operation	Before (ms)	After (ms)	Speedup	Serialized bytes
Cereal	1.5M IDs	write	5.789	0.391	14.81×	12,000,009
Cereal	1.5M IDs	read	10.377	0.379	27.37×	12,000,009
YAS	1.5M IDs	write	1.440	0.585	2.46×	12,000,015
YAS	1.5M IDs	read	7.109	0.205	34.75×	12,000,015
Boost binary	1.5M IDs	write	5.136	0.396	12.98×	12,000,008
Boost binary	1.5M IDs	read	9.882	0.388	25.45×	12,000,008

Run the benchmark:

./build/benchmarks/all/reflect-cpp-all-format-benchmarks \
  --benchmark_filter='bulk_numeric_vectors/[^/]+/ids/' \
  --benchmark_min_warmup_time=0.1 \
  --benchmark_min_time=0.2s \
  --benchmark_repetitions=5

[liuzicheng1987]

That's a great contribution. Thanks!