gh-153568: Use FIRST sets to speed up parser choices

[pablogsal]

The parser currently tries every alternative in a choice, even when the next token rules most of them out. This uses FIRST sets from the grammar to skip that work. When only one alternative can match, it jumps straight there. When several can match, a small bitmask keeps them in grammar order. The optimization leaves nullable choices alone and is disabled for the invalid-rule pass, so syntax errors stay the same.

On my machine, parsing every Python file in Lib dropped from 2.180 seconds to 1.868 seconds, about 14.6%. That costs roughly 45 KB more text in the parser extension, or 7.4%. Part of gh-153568.

[pablogsal]

I grabbed some perf counters while parsing every parseable Python file under Lib. I alternated the two builds on the same pinned CPU. main is 999a046 and the PR build is d9cca76.

Metric	main	this PR	Change
Parse time per Lib pass	2.215 s	1.891 s	-14.8%
CPU cycles	33.91 B	28.83 B	-14.8%
Instructions	106.15 B	91.88 B	-13.5%
Branches	25.36 B	21.86 B	-13.8%
Branch misses	82.05 M	80.85 M	-1.5%
Branch miss rate	0.324%	0.370%	+0.046 pp
Instructions per cycle	3.13	3.19	+1.8%