Implement tie_weights(token_embedding, lm_head_weight), which returns the weight matrix the LM head should use.
token_embedding is a numpy array of shape (vocab_size, hidden_dim); lm_head_weight is an array of the same shape or None.lm_head_weight is None, return token_embedding itself (the same array object, so edits to one are visible in the other).lm_head_weight unchanged if its shape matches token_embedding, and raise ValueError if it does not.