gh-153568: Reduce parser allocations

[pablogsal]

Allocate tokens in blocks and keep text only for names, literals and type comments. Use a bit for the Barry spelling and put the keyword error messages in the grammar. Create the identifier cache on first use and reuse the newline buffer. Reuse parsed names and numbers during backtracking, and reset memo filters before the error pass.

Related to #153568.

• Issue: Speed up the parser #153568

Codebase	Python files	Parsing speed	Peak parser memory
Standard library	2,060	8.0% faster	11.1% less
Home Assistant	18,731	4.4% faster	6.7% less
mypy	442	5.9% faster	10.7% less
PyTorch	4,890	7.2% faster	5.1% less

[pablogsal]

This looks gigantic because the parser autogenerated changes but it's quite tiny :)