Clustering : ReadData messages for ReadOnly Transactions sometimes end up in dead letters causing timeouts

From debugging it seems like this is happening because the ShardTransaction actor is killed before the ReadData message makes it to the actor. This appears to be related to our usage of PhantomReferences to track ReadOnly transactions. The reference sometimes seem to get collected right after the transaction is created and before the read data message could be send to the transaction resulting in the actor being killed and the read data being lost.

We do need to shut down read-only transaction actors ASAP as they could hold on to a lot of memory but we may need to review our usage of phantom reference and see if we can come up with a better alternative.