Reuse sharded shared batch buffers

[bjmeetsfo]

Summary

• add ShardedMultiLayerCache::get_shared_batch_into for caller-owned result buffers
• delegate colocated shared batches directly to the owning shard to avoid an extra merge allocation
• add regression coverage for colocated and cross-shard result-buffer reuse

Validation

• cargo fmt -- --check
• cargo check --locked --no-default-features --examples -q
• cargo test --locked --no-default-features "sharded_get_shared_batch" -- --test-threads=1
• cargo test --locked --no-default-features -q
• python3 -m py_compile tools/validate_backend_report.py tools/compare_backend_reports.py tools/emit_backend_operator_log.py tools/gen_metrics.py tools/validate_grafana_dashboard.py
• python3 tools/validate_grafana_dashboard.py
• git diff --check