flux_log() should do something sensible if flux_t argument is NULL

[garlick]

As noted in #321, it would make sense for flux_log() to log to stderr if the flux_t argument is NULL.

One use case (unit testability of otherwise standalone code with flux_log() calls in it) is mentioned in #1190

We could simply log to stderr if the arg is NULL. For extra credit, turn level into a string and log that too. I think it's debatable whether we'd want to go further than that, e.g. add timestamp, use globals to store the "log context" (thereby allowing flux_log_set_appnum() and flux_log_set_procid() to work), etc.