Fix: verify treats every Unbox as deopting (#271); two red develop tests - #290

• TypedPython: verifier and cgen disagree on whether an F64/BOOL Unbox can deopt #271: verify set can_deopt only for Unbox to i64, cgen treats every Unbox as deopting; verify now agrees (red: F64/BOOL Unbox accepted in an impure function). Hand-built fixtures that used an impure F64 Unbox use ObjToFloat. The front end emits no Unbox, so benchmarks are unaffected.
• test_verify_classes: the class-guard test still asserted the exemption removed by Fix: impure caller of a class-guarded callee is skipped, not a cgen crash (#275) #279 (red on develop since Fix: impure caller of a class-guarded callee is skipped, not a cgen crash (#275) #279).
• test_eval_breaker (Fix: chunked range loops poll once per chunk (#280) #284): the ratio bound 0.5 was too tight on CI (0.6 observed); 0.8 still fails the pre-TypedPython: chunked range loops poll every 16.7M iterations instead of every 4096 #280 code, which waited for the whole loop.

Closes #271