I have always done the second method, and just accepted the extra overhead of one extra message pass. Beware of premature optimization. It is generally rare for me to want to just forward a message like this without the forwarding actor needing to change or react to the message in some way.
An alternate design is to accept that you don't have an actor-subactor relationship, but that your subactor should really be a helper loop of the actor. A dedicated helper loop can share references no problem. Your "actors sharing references" is a potentially suboptimal mix of "actors are highly independent but follow restrictive rules" and "helper loops have no restrictions but are completely coupled to their owner"