Call convert_to_tensor in tf2jax.cast().

The intention is to get code like `tf2jax.cast(tfp.util.DeferredTensor)` to work correctly. As it is now, it calls `DeferredTensor.__array__`, which is only intended to work with numpy at the moment. It's not clear whether `jnp.array(DT)` should bypass numpy or not.

That aside, calling convert_to_tensor at the beginning of functions is pretty standard, so this change is probably a good idea either way.

PiperOrigin-RevId: 398321147

Author: SiegeLordEx

tensorflow_probability/python/internal/backend/numpy/ops.py

@@ -377,7 +377,7 @@ def batch_jacobian(self, target, source,  # pylint: disable=unused-argument
 
 
 def _cast(x, dtype):
-  x = convert_to_tensor(x)
+  x = np.asarray(x)
   if (np.issubdtype(x.dtype, np.complexfloating) and
       not np.issubdtype(dtype, np.complexfloating)):
     x = np.real(x)