Use native-backed AST nodes in the Rust parser

[adamziel]

What changed

• Return native-backed WP_MySQL_Native_Parser_Node wrappers from the Rust parser instead of eagerly materializing the whole AST into PHP objects.
• Delegate AST read methods to Rust-owned arena accessors, while preserving stable PHP node identity and materializing a node on first mutation.
• Speed up native lexing by replacing linear keyword/version/synonym scans with lazy static lookup maps and avoiding unnecessary identifier copies/uppercase allocation.
• Use a hybrid FIRST-set representation in the native parser.
• Replace the native AST wrapper cache HashMap with a dense Vec<Option<usize>> keyed by arena node index.
• Add --consume=none|descendants to run-parser-benchmark.php so lazy AST parser-only numbers can be compared with a forced full-tree traversal.

Benchmarks

Local PHP 8.4.5, 69,577 MySQL server-suite queries, 3 reps each.

benchmark	pure PHP avg	native avg	speedup
lexer-only, no JIT	107,823 QPS	1,139,760 QPS	10.6×
lexer-only, JIT	222,725 QPS	1,177,436 QPS	5.3×
parser-only, no JIT (--consume=none)	19,348 QPS	172,980 QPS	8.9×
parser-only, JIT (--consume=none)	41,706 QPS	181,311 QPS	4.3×
full AST walk, no JIT (--consume=descendants)	13,367 QPS	55,749 QPS	4.2×
full AST walk, JIT (--consume=descendants)	21,512 QPS	58,004 QPS	2.7×

--consume=descendants is a better benchmark than parser-only for lazy AST behavior because it forces PHP to touch all 4,786,376 descendants. It is still somewhat pessimistic for the SQLite translator because the translator usually reads selected subtrees rather than calling get_descendants() on the root.

Profiling notes

• After the keyword lookup changes, keyword scanning disappears from the parser-only sample; the remaining hot path is recursive parser control flow plus allocator churn.
• In the full AST walk sample, time shifts to PHP boundary costs: object creation, zend_update_property, property/hash lookups, zval arrays, and object destruction.
• The dense wrapper cache specifically targets the visible hashing/re-hashing cost during full AST walks.

Verification

• cargo fmt --check
• release build with the local PHP config
• php -l tests/tools/run-parser-benchmark.php
• php -l src/mysql/native/class-wp-mysql-native-parser-node.php
• php -d extension="$EXT" tests/tools/verify-native-parser-extension.php
• php -d extension="$EXT" vendor/bin/phpunit tests/mysql/native/WP_MySQL_Parser_Instanceof_Tests.php
• php -d extension="$EXT" vendor/bin/phpunit tests/mysql/WP_MySQL_Server_Suite_Parser_Tests.php
• php -d extension="$EXT" vendor/bin/phpunit tests/WP_SQLite_Driver_Translation_Tests.php

Caveat

This makes WP_Parser_Node non-final so the native facade can subclass it. That keeps the pure-PHP parser methods free of native checks, but may slightly reduce pure-PHP/JIT parser performance. A final-preserving native-handle design is possible, but would be more invasive.