JAX MLP activation list can be smaller than layer list (index misalignment)

Nobody has claimed this yet.

In jax_mlp.py, `self.layers` is built with one entry per layer, but `self.activation_funs` filters out `"linear"` activations.

https://github.com/lnccbrown/LANfactory/blob/9301939e8fed95ee966bf41fe71bb68bd620d7fd/src/lanfactory/trainers/jax_mlp.py#L97-L101

That means `self.activation_funs` can be shorter than `self.layers`.
Later, forward indexes `self.activation_funs[i]` using the layer index, which can raise `IndexError` or apply the wrong activation when `"linear"` appears before the final layer.

```python
def test_mlp_jax_with_linear_hidden_activation():
model = JaxMLP(
layer_sizes=[10, 10, 1],
activations=["tanh", "linear", "linear"],
train_output_type="logprob",
train=True,
)

rng = jax.random.PRNGKey(0)
test_input = jnp.ones((5, 6))
params = model.init(rng, test_input) # fails with IndexError in old pattern
output = model.apply(params, test_input)

assert output.shape == (5, 1)
```

Observed failure:
`IndexError: tuple index out of range` at jax_mlp.py

Expected:
Activation mapping should stay layer-aligned even when `"linear"` is used in hidden layers.

Start in src/lanfactory/trainers/jax_mlp.py at the activation-list construction around lines 97–101, then trace how forward indexes activations by layer. Use the provided test_mlp_jax_with_linear_hidden_activation case with the shown JAXMLP configuration. Done means linear hidden activations remain layer-aligned and the model initializes, applies, and produces output with shape (5, 1).

Written by the indexing model from the issue text.