Use fused attention for SmolLM3

[GargiGupta-io]

Description of the change

SmolLM3's attention layer always uses the eager einsum and softmax path even
when Keras provides a fused dot-product attention operation. This bypasses the
flash-attention path available on supported JAX and PyTorch configurations.

This change follows the existing Qwen3 and Phi3 pattern: use
ops.dot_product_attention() when fused_attention_op_available() is true,
normalize the attention mask for the fused operator, and retain the existing
implementation as the fallback.

Regression tests cover masked and unmasked dispatch, the eager fallback, and
numerical equivalence between both paths.

Reference

• Related to Update model implementations to use flash attention #2046
• Follows the fused-attention approach established in Use Flash Attention if available #2058

Colab Notebook

Not applicable. This change does not add a new API or model.

Validation

• TensorFlow: 6 passed, 2 skipped, 2 subtests passed
• JAX: 6 passed, 2 skipped, 2 subtests passed
• PyTorch: 6 passed, 2 skipped, 2 subtests passed
• Ruff lint and format checks
• The remaining macOS causal-LM test requiring tensorflow-text fails the
  same way on the unchanged master branch.

Checklist

• I have added all the necessary unit tests for my change.
• I have verified that my change does not break existing code and works with all backends (TensorFlow, JAX, and PyTorch).
• My PR is based on the latest changes of the main branch (if unsure, rebase the code).
• I have followed the Keras Hub Model contribution guidelines in making these changes.
• I have followed the Keras Hub API design guidelines in making these changes.
• I have signed the Contributor License Agreement.

[gemini-code-assist]

Code Review

This pull request integrates fused attention support into the SmolLM3Attention layer by utilizing ops.dot_product_attention when fused_attention_op_available() is true. It also adds comprehensive unit tests in smollm3_backbone_test.py to verify the fused attention execution, its fallback path, and numerical consistency between both implementations. There are no review comments to address, and the changes conform to the repository's guidelines.