fix(optuna): harden literal parsing across both desubstitution paths

[GrigoryEvko]

The AST-level _EvalCleaner (add_tuned_comment=False path) checks isinstance(arg, (Constant, List, Tuple, Set, Dict)) to decide whether an eval(...) call wraps a literal that can be inlined. This misses UnaryOp(USub, Constant), so eval(-5) survives in the AST path while the source-level _clean_eval_in_source strips it correctly via ast.literal_eval. The two desubstitution modes silently produced different output for the same input.

Reproducer against current main:

import ast
from gigaevo.programs.stages.optimization.optuna.desubstitution import (
    _EvalCleaner, _clean_eval_in_source,
)

src = "x = eval(-5)"
tree = ast.parse(src); _EvalCleaner().visit(tree); ast.fix_missing_locations(tree)
print(ast.unparse(tree))           # before: 'x = eval(-5)' ;  after: 'x = -5'
print(_clean_eval_in_source(src))  # always: 'x = -5'

Fix: in _EvalCleaner.visit_Call, replace the isinstance check with ast.literal_eval(arg) as a try/except predicate so both paths share the same literal-detection logic. Verified across 12 sample inputs — the AST and source paths now produce identical output. Also broaden the (ValueError, SyntaxError) catches around literal_eval in _coerce_param_value and _clean_eval_in_source to also catch MemoryError, RecursionError, and TypeError.