gh-153568: Reuse shared prefixes in generated parser rules - #157465

For obj[index], the parser tries the attribute and call alternatives before the subscript one. Each choice starts by asking for the same obj again. The parse cache saves us from parsing it from scratch, but we still pay for another rule call and cache lookup.

This teaches the generator to keep that result and the position where it ended in local variables. When one choice fails, the next one can pick them up directly and try its own ending. It does this for adjacent choices that start with the same memoized rule, including expression, arithmetic and assignment-target rules.

The grammar stays as it is. The generator finds these shared starts on its own.

On my machine, this makes parsing the standard library and twenty repositories about 10.7% faster overall.

The table compares this change against main at fd0970c0ab, using release builds without PGO or LTO. Each result is the median of five paired runs. The timings include tokenization, building the AST and cleanup, without compiling bytecode or running the files.

This overlaps with #157437, which shares some of these starts by rewriting the grammar. Adding this on top of that PR still gives about 5.2% faster parsing across the standard library, mypy, Home Assistant and PyTorch in a separate run. That comparison does not include the rest of the pending parser changes.

Part of #153568.

• Issue: Speed up the parser #153568