fix: Recurrent is actually a square matrix

Each neuron is affected by all states.

With this change the forward procedure is working correctly. I verified
a couple of test cases.

Author: castelao

src/nf/nf_rnn_layer.f90

@@ -20,7 +20,7 @@ module nf_rnn_layer
     integer :: output_size
 
     real, allocatable :: weights(:,:)
-    real, allocatable :: recurrent(:)
+    real, allocatable :: recurrent(:,:)
     real, allocatable :: biases(:)
     real, allocatable :: z(:) ! matmul(x, w) + b
     real, allocatable :: state(:)

src/nf/nf_rnn_layer_submodule.f90

@@ -137,7 +137,7 @@ module subroutine init(self, input_shape)
     ! Broadcast weights to all other images, if any.
     call co_broadcast(self % weights, 1)
 
-    allocate(self % recurrent(self % output_size))
+    allocate(self % recurrent(self % output_size, self % output_size))
     call random_normal(self % recurrent)
     self % recurrent = self % recurrent / self % input_size