Refactor per-query ANN I/O metrics to the object-store level by reusing IOTrackingStore

Context

PR #7204 (closes #7201) made per-query index I/O observable on the ANNSubIndex and ANNIvfPartition operators in EXPLAIN ANALYZE. Because the ANN operators delegate reads to a cached, shared IVFIndex whose ScanScheduler is reused across queries, the established "own a fresh scheduler and read its cumulative stats" pattern did not apply. The PR introduced a lightweight per-query sink, IoStatsRecorder (a new trait in lance-core) backed by lance_io::scheduler::IoStats, attached at the FileScheduler accounting point.

In the review of that PR (https://github.com/lance-format/lance/pull/7204#pullrequestreview-4470060186), @wjones127 approved the approach as acceptable for now but asked for a future refactor that moves this tracking to the object-store level and reuses the existing IOTrackingStore / IOTracker rather than the scheduler-level hook.

Proposed direction

From the review:

Perhaps IOTrackingStore could hold a Vec<Arc<IOTracker>> and ObjectStore could have some method like:

impl ObjectStore {
    fn with_io_tracker(tracker: Arc<IOTracker>) -> Self { ... }
}

So you can always wrap that.

IOTracker already wraps an object store and accumulates IoStats (rust/lance-io/src/utils/tracking_store.rs). The idea is to let a caller attach a per-scope IOTracker to the object store an operator reads through, so per-query metrics fall out of the existing object-store tracking machinery instead of a parallel scheduler-level sink.

Acceptance criteria / open questions

• Per-query bytes_read / iops / requests on ANNSubIndex and ANNIvfPartition stay exact and ~0 on a warm cache (parity with the tests added in #7204).
• Metrics remain attributed per query, not process-wide cumulative, despite the shared/cached IVFIndex scheduler.
• Evaluate whether IoStatsRecorder (lance-core) and the scheduler-level hook can be removed once the object-store path covers this, or whether both must coexist.
• Confirm the object-store-level counts match what the FileScheduler accounting point reports today (post-coalescing physical I/O).

Follow-up to #7204. Optional / non-blocking refactor.

Start with rust/lance-io/src/utils/tracking_store.rs and the existing IOTracker/IOTrackingStore path, then inspect IoStatsRecorder in lance-core and the ANNSubIndex and ANNIvfPartition metrics from #7204. Run the per-query ANN I/O tests added in #7204 as a baseline. Done means object-store tracking preserves exact per-query bytes_read, iops, and requests, remains near zero on warm cache, and the scheduler hook is removed or its coexistence is justified.