Expose SessionContext.add_optimizer_rule for Python-defined logical optimizer rules

まず、SessionContext.remove_optimizer_rule の既存の Python binding と、PR #1557 の add_physical_optimizer_rule の作業を確認してください。upstream で FFI_OptimizerRule と Python LogicalPlan コンストラクターが提供されたら、この issue を再度確認してください。Python で定義した論理ルールを add_optimizer_rule を通じて登録し、binding を通じて実行できれば完了です。

Background

SessionContext::add_optimizer_rule registers an Arc<dyn OptimizerRule> that runs during logical plan optimization. The Python bindings only expose remove_optimizer_rule, never the additive side. PR #1557 ultimately landed add_physical_optimizer_rule for the physical pipeline via FFI but did not address the logical pipeline.

Upstream signature

pub fn add_optimizer_rule(&self, optimizer_rule: Arc<dyn OptimizerRule + Send + Sync>)

User value

Lets users add domain-specific logical rewrites (predicate normalization, redundant join elimination, scan pruning informed by external metadata) without forking DataFusion. Complements the existing remove_optimizer_rule to round out the surface.

Why deferred

Blocked upstream. As documented in PR #1557, DataFusion does not currently expose an FFI bridge for the logical OptimizerRule / AnalyzerRule traits, and there are no Python constructors for LogicalPlan node variants -- a pure-Python rule could observe plans but not transform them. The PR #1557 commit history shows an initial attempt at a Python-defined logical rule that was abandoned for this reason. This is filed for tracking; it should be revisited once upstream lands an FFI_OptimizerRule (mirroring FFI_PhysicalOptimizerRule).