[X-2692] Vortex Segment Cache

[ch-sc]

Summary

VortexSource had no way to share segment bytes across queries. Every DataFusion query re-read the same zone-map and data segments from object storage, even though for an unchanged file those bytes are identical. This change adds a cross-query, cross-file segment cache, plumbed end-to-end from VortexSource down to VortexOpenOptions.

The naive approach of passing one Arc<dyn SegmentCache> for all files is unsafe: SegmentId is a file-local u32 index, e.g., segment 5 from file A and segment 5 from file B would alias to the same cache entry and silently serve wrong data.

This PR introduces a SegmentCacheBuilder indirection that namespaces per-file caches by a FileIdentity (path + content version) so a single shared cache is correct across many files.

The SegmentCacheBuilder builder hands out per-file SegmentCache views from one shared backing store, namespaced by FileIdentity (path + etag/size-mtime). VortexSource::with_segment_cache_builder(...) wires it through. Build one NamespacedMokaSegmentCacheBuilder per process, reuse for every query — file-safe, overwrite-safe, cross-query reuse.

The InstrumentedSegmentCacheBuilder decorator SegmentCacheBuilder — it wraps another builder and adds hit/miss/store metrics to every per-file cache that builder produces.

Usage

  use std::sync::Arc;
  use vortex::layout::segments::{
      InstrumentedSegmentCacheBuilder, NamespacedMokaSegmentCacheBuilder,
  };
  use vortex_datafusion::VortexSource;

  // Built once per process / SessionContext:
  let cache = Arc::new(InstrumentedSegmentCacheBuilder::new(
      NamespacedMokaSegmentCacheBuilder::new(2 << 30), // 2 GB shared budget
      metrics_registry,
      vec![],
  ));

  // Reused for every VortexSource:
  let source = VortexSource::new(schema, session)
      .with_segment_cache_builder(Arc::clone(&cache));

[ch-sc]

Closing this as we need to merge into vortex 0.65.0 and this one is way ahead. Switched to new PR: #5