Linking upstream DataFusion and arrow-rs source

• Clones live next to the main datafusion-distributed checkout: $UPSTREAM/datafusion and $UPSTREAM/arrow-rs, where UPSTREAM=$(cd "$(git rev-parse --path-format=absolute --git-common-dir)/../.." && pwd). The user may have work in progress in them, so link a detached worktree named after the version instead ($UPSTREAM/arrow-rs-<version>), e.g. git -C "$UPSTREAM/arrow-rs" worktree add --detach "$UPSTREAM/arrow-rs-<version>" '<version>^{commit}'.
• The versions in use are in Cargo.lock. These crates are published from the voted release tarball, so they ship no .cargo_vcs_info.json. The release tag (<version>, no v prefix) is the last RC's commit, which is the published source. Check a crate with diff -r ~/.cargo/registry/src/*/<crate>-<version> <worktree crate dir>.
• Patch releases come from maintenance branches (branch-55, 59_maintenance), so fetch the tag explicitly: git fetch <apache remote> tag <version>.
• Link with [patch.crates-io] entries in an untracked .cargo/config.toml, using absolute paths. Every crate of the family that appears in Cargo.lock must be patched. Patching only arrow-ipc builds a second arrow-array from the worktree, and its types don't unify with the crates.io copy. cargo metadata --no-deps --format-version 1 in the worktree maps crate names to their directories.
• A patch applies only to the locked version. When the lock pins another compatible version (e.g. linking arrow 59.3.0 into a DataFusion checkout locked to 59.2.0), cargo silently keeps crates.io until you run cargo update -p <crate> for each patched crate. Always confirm with cargo tree -i <crate> --depth 0 before measuring.
• Linking rewrites Cargo.lock. Never commit it or .cargo/ (some IDEs stage new files automatically). Unlink with rm -rf .cargo && git checkout Cargo.lock. Linking and unlinking each trigger a full rebuild.
• Worktree edits apply on the next build. To keep a fix, git switch -c in the worktree. Upstream PRs target main, so port the change there. Upstream's own tests need git submodule update --init.
• Every upstream fix or performance claim must be reproducible in the upstream repo alone. Upstream reviewers won't run datafusion-distributed, so its results only motivate the change. Demonstrate it with an upstream test, or with the upstream benchmarks: criterion benches in each crate (cargo bench -p <crate> --bench <name> -- --save-baseline before, then --baseline before) and DataFusion's benchmarks/bench.sh.
• Existing upstream benches may not exercise the conditions where the problem shows up here, such as many concurrent tasks or larger batches. Single-threaded benches miss allocator contention, for example. Extend them until the effect reproduces, and report any cases that regress.