Don't swallow errors with RNG keys in parameter initialisers.

Given the following code snippet:

```python
def f(x):
  x = hk.Conv2D(3, 1)(x)
  return x

f = hk.transform(f)
x = jnp.ones([1, 28, 28])
f.init(None, x)
```

At HEAD we fail with a confusing error:

    ValueError: Parameters cannot be `None`.

After this change we fail with a far more useful:

    MissingRNGError: You must pass a non-None PRNGKey to init and/or apply if you
    make use of random numbers.

PiperOrigin-RevId: 551458310

haiku/_src/base.py

@@ -683,6 +683,8 @@ def get_parameter(
     except MissingRNGError as e:
       if frame.params_frozen:
         raise param_missing_error from e
+      else:
+        raise e
 
     if param is DO_NOT_STORE:
       # Initializers or custom creators that return `DO_NOT_STORE` are required

haiku/_src/base_test.py

@@ -137,6 +137,15 @@ def test_parameter_in_immutable_ctx(self, params):
           ValueError, "parameters must be created as part of `init`"):
         base.get_parameter("w", [], init=jnp.zeros)
 
+  def test_get_parameter_rng_exception(self):
+    with base.new_context():
+      with self.assertRaisesRegex(
+          base.MissingRNGError, "pass a non-None PRNGKey to init"
+      ):
+        base.get_parameter(
+            "w", [], init=lambda shape, dtype: base.next_rng_key()
+        )
+
   def test_get_parameter_wrong_shape(self):
     with base.new_context():
       with self.assertRaisesRegex(ValueError, "does not match shape"):