Replicate results instead of commands in `MODELRUN`

[lantiga]

Instead of replicating MODELRUN verbatim, we should just replicate the result.

This amounts to using RedisModule_Replicate and sending AI.TENSORSET instead of AI.MODELRUN, with the serialized output tensors as arguments.

This needs to happen in RedisAI_Run_Reply, once the computation has finished and the client has been unblocked and the response is being sent (since that is the first opportunity to have the outputs available in the main thread)

https://github.com/RedisAI/RedisAI/blob/master/src/redisai.c#L637

[lantiga]

Two comments:

1. the same holds for SCRIPTRUN
2. with reference to Make OUTPUTS optional in MODELRUN and SCRIPTRUN #86, MODELRUN and SCRIPTRUN should only be replicated if output keys are specified, otherwise they should not (since the variant that returns the outputs directly does not have any side-effect on the server side) (not actual, see Add DAGRUN (or PIPERUN) command #88)

[lantiga]

Resolved by #157