Better IO orchestration

[robert3005]

Vortex reader will collect all read requests from layouts and dispatch them together https://github.com/spiraldb/vortex/blob/develop/vortex-serde/src/layouts/read/stream.rs#L192. However, this is extremely naive and doesn't leverage additional knoweldge we have about file format to prioritize requests and prefetch data.

AnyBlob outlines techniques for improving throughput and latency on blob stores. We should explore methods outlined in the paper and try to incorporate them into vortex file reading.

In no particular order the things to look at are

• Better queuing, we shouldn't wait for arbitrary chunk to finish reading if we have chunks we can process already
• Performing io via io_uring (on linux) and minimise copies
• On disk caching to allow deduplication across files. We could checksum vortex batches and deduplicate loading across files using on disk cache

[a10y]

Fusio is almost a good fit for us as a replacement for object_store that is runtime agnostic, but it's missing two things

1. Support for a ReadAt trait
2. An implementation of the S3 client that runs on monoio

I'm working with fusio and monoio folks on both of these, since monoio affects it's implementation for (1).

Monoio: https://github.com/bytedance/monoio/pull/309

Fusio: https://github.com/tonbo-io/fusio/issues/68

[ethe]

I implement read_exact_at in fusio: https://github.com/tonbo-io/fusio/pull/71/files#diff-22e7f097629cff0a2a265b5185882031f09033e6e67d8bf2303e2c37314b1a7fL50 , please take a look