feat: add GPU buffer loader for IndexProvider integration

[cluster2600]

Summary

• GpuBufferLoader (gpu_buffer_loader.h): streams vectors from any IndexProvider into contiguous GPU-ready float32 buffers
• Metal C++ docs (docs/METAL_CPP.md): architecture overview and kernel reference

Replaces #174 (now closed), which incorrectly used a standalone RocksDB store. This PR integrates with zvec's existing storage architecture via IndexProvider::Iterator.

Follow-up to #166 ("Future Work: Integration with storage").

How it works

IndexProvider (Flat/HNSW/IVF)
    |
    +-- Iterator -> GpuBufferLoader::load() -> GpuBuffer
                                                  |
                                        +---------+----------+
                                        |                    |
                                  Metal device buf      cudaMemcpy

auto provider = index->create_provider();
auto buffer = zvec::GpuBufferLoader::load(provider);

// buffer.vectors is contiguous (N x dim) float32
// Ready for Metal newBufferWithBytes or cudaMemcpy

Features

• load() — stream all vectors into a single contiguous buffer
• load_chunk() — chunked loading for datasets exceeding GPU memory
• Automatic type conversion — FP16, INT8 -> FP32
• Works with all index types — Flat, HNSW, IVF providers

Why not RocksDB?

zvec already has a complete storage stack: IndexProvider -> Iterator -> block-based segments with mmap/buffer pool backends. A parallel RocksDB store would duplicate this. GpuBufferLoader sits on top of the existing pipeline instead.

Merge order

This PR shares a common base with #172, #173, #176. Recommended merge order: #172 → #173 → #175 → #176. Merging any one brings in the shared base commits; the rest then apply cleanly.

Test plan

• Header compiles with clang++ C++17
• Integrates with existing IndexProvider / IndexHolder::Iterator interfaces
• End-to-end: Flat provider -> GpuBufferLoader -> Metal compute pipeline
• Benchmark: load throughput for 1M+ vectors
• Test FP16 and INT8 conversion paths