Safe casting of PgNode values with bindgen tag discovery

[cbandy]

Postgres implements various node types to represent parse, plan, and executor trees. It uses tagged polymorphism to emulate object-oriented inheritance in C: every node struct begins with a NodeTag field that identifies its type at runtime. Pointers to these nodes are accessed generically as Node * and queried using the IsA macro. The hierarchy includes intermediate base types, such as Expr, which are inherited by concrete types like Const and Var.

This PR introduces type-safe node casting (both upcasting and downcasting) via the PgNode trait. Safe casts are statically generated with the trait implementations emitted by pgrx-bindgen. Each node type may be cast to any of its "parent" types, all the way up to Node.

The PgNode trait is extended with:

• A constant CAST_TAGS slice containing the valid tags for the type and all of its descendants. This is efficient because the hierarchy is shallow and wide. The vast majority of types have four or fewer tags in CAST_TAGS.
• A try_cast associated function and a try_as method check CAST_TAGS before performing an unsafe pointer cast. The lookup is a simple linear scan and relies on compiler and CPU optimizations.
• A node_tag method that returns the tag and an is_a method that behaves like the IsA macro.

See: #1048

[cbandy]

The first commit adds to impl_pg_node in pgrx-bindgen, mostly independently of the existing implementation. The second commit is more invasive and rearranges the function to find nodes and accumulate tags simultaneously. Both perform similarly, so we should use the one that seems easier to read/maintain.

I had an agent compare all three commits using hyperfine:

Benchmark Results: Postgres Node Hierarchy Traversal

Below is the comparative performance benchmark measuring compilation and binding generation for pgrx-pg-sys with the release-time code path:

PGRX_PG_SYS_GENERATE_BINDINGS_FOR_RELEASE=1 cargo build -p pgrx-pg-sys --no-default-features -F pg18

Metric	develop (No Feature)	HEAD^ (Original Cast Feature)	HEAD (Refactored Cast Feature)
All Runs (Mean)	16.867 s ± 1.455 s	17.260 s ± 1.593 s	17.155 s ± 1.627 s
Warm Runs (Mean) ¹	16.221 s	16.551 s	16.431 s
Minimum Build Time	16.030 s	16.425 s	16.287 s
Max Memory Usage	1.72 GB (1,716,551,680 B)	1.78 GB (1,784,430,592 B)	1.74 GB (1,740,562,432 B)

¹ Warm Runs Mean ignores the first run to account for filesystem cache warm-up.

Key Takeaways

• Build Time Overhead:
  • The original implementation (HEAD^) added ~330ms to build times (a ~2.0% compile-time increase over develop) because of complex recursive DFS traversals and syntax item filtering.
  • The refactored version (HEAD) reduces the warm-build mean to 16.431s, recovering ~120ms (or ~36%) of the added overhead.
• Memory Usage Overhead:
  • The original implementation (HEAD^) added ~68 MB of memory overhead (a ~4.0% increase over develop) during compilation.
  • The refactoring in HEAD (which eliminates the intermediate node_set and node_map structures and optimizes name mapping lookups) brings the memory overhead down to ~24 MB over develop, reclaiming ~44 MB (or ~65%) of the memory overhead introduced in HEAD^.

[cbandy]

1. These files generated quite differently on my system, so I only included the changes to PgNode impl here.
2. I also don't have PG 19 set up yet, so that is missing from the PR.

Should this PR include these changes, or perhaps are they regenerated only when it is time to tag?