fix: Replace VLAs with std::vector for MSVC compatibility and stack safety

[feihongxu0824]

No description provided.

[greptile-apps]

Greptile Summary

This PR systematically replaces Variable Length Arrays (VLAs) with std::vector across the entire codebase to achieve MSVC compatibility and improve stack safety.

Key Changes:

• Replaced all VLA declarations (e.g., float arr[size]) with std::vector<T> (e.g., std::vector<float> arr(size))
• Added #include <vector> to files that needed it
• Updated function calls to use .data() when passing vectors to functions expecting pointers
• Corrected sizeof() usage to properly calculate byte sizes for vectors (e.g., sizeof(T) * vec.size())
• In test files, replaced some memset calls with more idiomatic std::fill

Impact:

• MSVC Compatibility: VLAs are a C99 feature not supported by MSVC and not part of the C++ standard
• Stack Safety: Large VLAs can cause stack overflow; heap-allocated vectors are safer
• Standards Compliance: Makes the codebase fully C++ standard compliant

Performance Considerations:
In performance-critical SIMD code paths (inner_product_distance_batch_impl*.h), vectors are allocated inside loops. While this introduces heap allocation overhead, the impact should be minimal because:

• Template parameters like dp_batch are typically small compile-time constants
• Modern allocators are efficient for small allocations
• The compiler may optimize some allocations
• The benefits (portability, safety, standards compliance) outweigh the small performance cost

All changes follow a consistent mechanical pattern and appear to be correctly implemented.

Confidence Score: 4/5

• This PR is safe to merge with low risk - changes are mechanical and consistent across all files
• Score of 4 reflects high confidence with minor reservation: while all changes are mechanically correct and improve portability/safety, there are vector allocations inside performance-critical loops that could have small runtime impact. The extensive scope (38 files) requires thorough testing to ensure no edge cases were missed.
• Pay attention to src/ailego/math_batch/inner_product_distance_batch_impl*.h files - these contain SIMD hot paths with loop-scoped vector allocations that should be performance tested

Important Files Changed

Filename	Overview
src/ailego/algorithm/lloyd_cluster.h	Replaced VLAs with std::vector for scores, nearest_scores, and nearest_indexes arrays; added vector include; all .data() conversions correct
src/ailego/math/mips_euclidean_distance_matrix.h	Replaced all fixed-size u2 and v2 arrays with std::vector across multiple template specializations; consistent pattern applied
src/ailego/math_batch/inner_product_distance_batch_impl.h	Replaced VLAs for SIMD registers (accs, data_regs, sum128_regs) with std::vector; some allocations occur inside loops which may have minor performance impact
src/ailego/math_batch/inner_product_distance_batch_impl_fp16.h	Replaced VLAs for FP16 SIMD registers with std::vector across multiple implementations (AVX512FP16, AVX512F, AVX2); loop-scoped allocations present
src/core/algorithm/hnsw/hnsw_algorithm.cc	Replaced neighbor_ids, dists, and neighbor_vecs VLAs with std::vector; added vector include; proper .data() usage throughout
src/core/algorithm/hnsw/hnsw_entity.cc	Replaced padding, mapping, and buffer arrays with std::vector; correctly updated sizeof() calculations to use vector.size() * sizeof(element)
src/core/algorithm/hnsw_sparse/hnsw_sparse_algorithm.cc	Replaced neighbor_ids VLAs with std::vector in sparse HNSW implementation; added vector include; proper .data() usage
src/core/algorithm/ivf/ivf_entity.cc	Replaced distances VLA with std::vector in IVF search functions; proper .data() usage for distance calculations

_{Last reviewed commit: f8bab86}

[feihongxu0824]

@greptile

[iaojnh]

@greptile

[iaojnh]

lgtm