Proposal: GPU buffer loader for IndexProvider integration

[cluster2600]

Motivation

GPU backends (FAISS GPU, cuVS, Metal) need contiguous float32 buffers, but zvec stores vectors in IndexProvider segments. Currently there's no bridge to stream vectors from storage into GPU-ready memory.

Proposed approach

A new C++ class GpuBufferLoader (src/ailego/gpu/gpu_buffer_loader.h) that:

1. Iterates over an IndexProvider via its existing Iterator API
2. Deserializes ForwardData into float32 vectors
3. Fills a contiguous host buffer suitable for GPU upload
4. Reports progress (vector count, bytes loaded)

IndexProvider → Iterator → ForwardData → float32 buffer → GPU

This stays within zvec's existing storage architecture — no standalone databases or new storage engines.

Also includes a docs/METAL_CPP.md documenting the Metal kernel architecture and GPU integration points.

Questions for maintainers

1. Is IndexProvider::Iterator the right abstraction for bulk vector extraction, or is there a better path?
2. Should the buffer loader live under src/ailego/gpu/ or somewhere else?
3. Any concerns about memory usage for large collections? (Current approach loads everything into host RAM before GPU upload — we could add batched streaming.)

Draft implementation: #175

[richyreachy]

That's right. Index provider can be used to access vector data within an index. As for gpu buffer loader, let's take it into an overall GPU related design when it's time to start. Thanks very much for submitting PR. The effort is really appreciated.

[richyreachy]

I close this for now. We can refer back to it once the related issue is on the table. Thanks.