Add optional liquid cache for indexed parquet scans

[alchemist51]

What

Adds an opt-in in-memory decoded-batch cache (liquid cache) for the DataFusion analytics backend, wired behind a cargo feature so it is completely absent from the default build. Liquid cache lives outside this repo — it's pulled in only when a user explicitly points the build at a liquid cache source.

Behaviour

• Feature off (default): the integration is an inlined no-op — eligible scans use the plain ParquetSource unchanged. No liquid cache dependency is declared, resolved, fetched, or compiled. The build behaves exactly as before.
• Feature on: an eligible indexed row-group scan — all projected columns numeric/date/timestamp/boolean, within a column budget, and no string/binary predicate column — is wrapped with LiquidParquetSource, so decoded batches are served from and populated into the process-global cache.

How to enable

Two steps — a single-file edit plus a build flag:

1. In sandbox/plugins/analytics-backend-datafusion/rust/Cargo.toml, uncomment the liquid cache dependency lines, set the git repo + branch to your liquid cache source, and uncomment the matching dep: entries on the liquid_cache feature (all three edits are adjacent, with inline instructions).
2. Build with -PliquidCache (or LIQUID_CACHE=1).

./gradlew ... -PliquidCache

For local development against a checkout, redirect the git dep with a workspace [patch].

Why a commented placeholder instead of a committed dependency

Cargo resolves (and fetches) every declared optional dependency even when its feature is off — only compilation is feature-gated. A committed git URL would therefore be fetched on every default build, and a placeholder/unreachable URL would break it (an invalid URL even fails manifest parsing). Leaving the dependency commented out is what keeps the default build free of any liquid cache source while still making enablement a one-file edit.

Changes

6 files, additive (Cargo.lock intentionally not committed; the repo's Rust build runs unlocked and regenerates it):

• analytics-backend-datafusion/rust/src/liquid_cache.rs — feature-gated module. maybe_wrap_parquet_source is a no-op when off; when on, applies the eligibility gate and wraps the source. Holds the process-global cache reference.
• analytics-backend-datafusion/rust/src/indexed_table/parquet_bridge.rs — 12-line hunk routing the parquet source through maybe_wrap_parquet_source.
• analytics-backend-datafusion/rust/Cargo.toml — commented placeholder dependency + liquid_cache feature; all liquid-cache config consolidated here.
• dataformat-native/rust/Cargo.toml — pointer note (config lives in the plugin crate).
• dataformat-native/build.gradle — -PliquidCache / LIQUID_CACHE=1 toggles the cargo feature.
• analytics-backend-datafusion/rust/src/lib.rs — module declaration.

Testing

• Feature off: cargo check -p opensearch-datafusion clean, cargo fmt --check clean, no liquid cache source resolved.
• Enable path: simulated the one-file edit against a real repo/branch — cargo metadata resolves and pulls the liquid cache crates.

Notes / follow-ups

• Feature-on currently compiles only on Linux (the upstream liquid cache disk tier uses io-uring).
• The runtime bootstrap that populates the cache reference is not wired yet; the wrap is inert until that lands.