Fix accuracy issues with Gemma models

[kunal-vaishnavi]

Description

This PR fixes accuracy issues with Google's Gemma models by using bfloat16 precision, always using float32 precision to compute any LayerNorms, and casting the output logits to float32 always.

Motivation and Context

This PR has been tested with Gemma-2 and Gemma-3. It is using the bfloat16 changes from this PR as well as the missing final norm changes from this PR.

To fix the issue, we should avoid calling a potentially overridden method (make_outputs_init) in the __init__ method of the Model class. Instead, we can:

1. Rename the make_outputs_init method in the Model class to a private method (e.g., _make_outputs_init) to ensure it is not overridden by subclasses.
2. Call this renamed method (_make_outputs_init) in the Model class's __init__ method.
3. Subclasses like Gemma2Model can still define their own make_outputs_init method if needed, but it will not interfere with the Model class's initialization process.

This approach ensures that the Model class's initialization is self-contained and does not depend on subclass behavior.

---

To fix the issue, we will add an explicit return None statement at the end of the get_shape_of_value_info function. This ensures that the function always returns a value explicitly, making its behavior consistent and easier to understand. The change will not alter the existing functionality but will make the function's intent clearer.

---