Stop zero initializing the SAC networks

This makes gradient decent super hard, and it fails to converge.  The
default random initialization is much better.

Change-Id: I96465dad8313e5101a58365fee4458ffe7336442
Signed-off-by: Austin Schuh <austin.linux@gmail.com>
diff --git a/frc971/control_loops/swerve/velocity_controller/model.py b/frc971/control_loops/swerve/velocity_controller/model.py
index 0d8a410..1394463 100644
--- a/frc971/control_loops/swerve/velocity_controller/model.py
+++ b/frc971/control_loops/swerve/velocity_controller/model.py
@@ -113,7 +113,6 @@
         mu = nn.Dense(
             features=self.action_space,
             name='mu',
-            kernel_init=nn.initializers.zeros,
         )(x)
 
         log_std_layer = nn.Dense(features=self.action_space,
@@ -180,8 +179,10 @@
             )(x)
             x = self.activation(x)
 
-        x = nn.Dense(name=f'q', features=1,
-                     kernel_init=nn.initializers.zeros)(x)
+        x = nn.Dense(
+            name=f'q',
+            features=1,
+        )(x)
 
         return x