[FSDP] Don't use detach in _rebuild_full_params (#4813)

Summary:
Somehow the detach_copy implementation has some issues with FSDP that it messed up with FSDP's memory management mechanism. Let's not use it.

Test Plan:
CI

Author: alanwaketan

torch_xla/distributed/fsdp/xla_fully_sharded_data_parallel.py

@@ -1398,7 +1398,7 @@ def _rebuild_full_params(self,
 
     for p, p_shard in zip(self.full_params, self.sharded_params):
       if not p._has_full_param:
-        p_shard_data = p_shard.detach()
+        p_shard_data = p_shard
         if apply_opt_barrier:
           self.optimization_barrier_op([p_shard_data])
         if p_shard_data.dtype != self.compute_dtype: