Better hyperparams for GridWorld/SAC (#2776)

config/sac_trainer_config.yaml

     normalize: false
     num_layers: 1
     hidden_units: 128
-    init_entcoef: 0.01
+    init_entcoef: 0.5
+    buffer_init_steps: 1000
     buffer_size: 50000
     max_steps: 5.0e5
     summary_freq: 2000