fix(complexity): count what the visitor claims to count

[GrigoryEvko]

NumericalComplexityVisitor had three correctness gaps: (a) total_nodes was a sum of 7 specific visitor counts and under-reported by ~6× — use len(ast.walk(tree)) instead; (b) visit_Match and visit_AsyncFor were missing, so programs using match or async for reported condition_count=0 / loop_count=0; (c) comprehensions, IfExp, Lambda, try/except, and except* were not counted at all.

Reproducer against current main:

from gigaevo.programs.stages.complexity import compute_numerical_complexity

print(compute_numerical_complexity("def f(): x=1+2; return x")["total_nodes"])
# before: 2;  after: 13

r = compute_numerical_complexity("[v for v in xs if v > 0]")
print(r["loop_count"], r["condition_count"])
# before: 0 0;  after: 1 1

Fix: use len(ast.walk(tree)) for total_nodes; add visit_Match (each case is a condition), visit_AsyncFor (loop), comprehension visitors (each generator = loop, each if clause = condition), visit_IfExp / visit_Lambda / visit_Try / visit_TryStar. Verified against an independent ast.walk-based reference on 9 small samples and 7 large production files from torch + transformers (4k–10k LOC); all five counts match across both implementations. One pre-existing test asserted the buggy sum-of-counts as expected output and is updated.