Agent (Player) unwrapped from the _FabricModule does not consider the chosen precision during env interaction

[belerico]

Every time, during the environment interaction, the we call agent.module to unwrap the agent from the distributed strategy, we also unwrap the agent from the precision plugin, this means that if we are training an agent with float16 or bfloat16 then the environment interaction happens in float32.

I suggest to wrap every player agent with a _FabricModule, i.e. _FabricModule(agent, precision=fabric.precision) so to unwrap the agent from the strategy but maintaining the precision plugin.

cc @michele-milesi