gh-153568: Factor shared prefixes in expression rules

[pablogsal]

Several expression rules start by parsing the same thing for each choice. For a - b, the parser tries the + alternative first, then goes back and gets a again to try -. The cache saves us from parsing a from scratch, but we still pay for the extra rule call and cache lookup.

This puts the shared part before the choice: get the left operand once, then check which operator follows. It does the same for tuples, boolean expressions, comparisons and powers. Parse the first expression once, then look for the optional part that follows it. When there is nothing more to match, return the expression we already have.

On my machine, this makes parsing the full standard library and ten repositories about 4.7% faster overall.

Codebase	Python files	Parsing speed
Standard library	2,060	4.8% faster
Django	2,932	5.5% faster
pandas	1,527	4.8% faster
NumPy	495	4.4% faster
SciPy	999	4.6% faster
scikit-learn	1,036	4.7% faster
SymPy	1,607	5.6% faster
Transformers	4,884	3.9% faster
PyTorch	4,890	4.4% faster
Home Assistant	18,731	4.7% faster
mypy	442	5.5% faster

The table compares this change against main at a60343ed17, using release builds without PGO or LTO. The timings include tokenization, building the AST and cleanup, without compiling bytecode or running the files.

Part of #153568.

• Issue: Speed up the parser #153568