gh-153568: Skip the expression precedence chain for single-token atoms

[pablogsal]

Parsing a bare name or number descends the whole expression precedence chain (about 14 rule invocations plus their memoization traffic) just to wrap one token in a node, and that shape dominates real code through call arguments, assignments and subscripts. When the next token cannot start or continue an expression the result is provably that single atom, so the chain entry rules now build it directly, leaving the token fill state exactly as before so error locations are unaffected.

Rules declare the hook in the grammar with a new (fastpath=function) flag next to (memo), so the generator only emits the hook call and knows nothing about tokens or expressions; the atom logic lives with the other token helpers in pegen.c, and the mechanism is documented with a worked example in InternalDocs/parser.md. Candidate hooks on more chain rules, assignment targets and subscript slices were measured individually and dropped: none cleared 0.4%, and hooking every chain level measured 6% slower from the per-entry decline checks.

Benchmark (parsing 8 of the largest stdlib files, 1.3 MB, 20 times per run with _PyParser_ASTFromString — parser only, no AST-to-Python conversion; pyperf, interleaved runs):

build	time per run	speedup
main	1.69 s ± 0.03
this PR	1.52 s ± 0.02	1.11x faster

• Issue: Speed up the parser #153568