Object endpoints: measuring after every call costs CPU; module endpoints and 401 headers left open #32

Follow-ups from the adversarial review of #31 (fix for #30), deferred there:

• Measuring costs time on every call. Class and instance endpoints walk the object, up to max_obj_size bytes of it, after every dispatched call, reads included. A read on a 10k-entry dict (575 KB) costs about 15 ms. One option is an obj_mutators= handler option that limits measuring to the names declared as mutating, while keeping the current "measure everything" as the default. Another is to throttle measuring.
• Lazy results are measured too early. A method that returns a generator or a coroutine does its work after the measure, so the overshoot is only caught by the next call. That is documented in Instance endpoints need an obj_owner and are bounded in size (#30) #31, not fixed.
• Module endpoints. A module exposed through attr_names has no owner hook and no size bound. Its functions can mutate module globals. The same is true of a bound method passed straight into an app spec.
• 401 responses from bottle_error_handler carry no WWW-Authenticate header (RFC 7235). Which scheme to name depends on the deployment's auth (JWT middleware, obj_owner).