[XLM] Refactor output tracing to align with capture_outputs standardized architecture

[aman-coder03]

What does this PR do?

This PR refactors XLM's output tracing to align with the standardized output capturing patterns used across the codebase.

Key changes:

• Refactors transformer blocks into a dedicated XLMLayer module to enable structured output capture
• Integrates the capture_outputs decorator with _can_record_outputs to automatically capture attention weights from each transformer layer
• Removes manual attention propagation logic and relies on the standardized output capturing infrastructure
• Implements explicit hidden_states collection in XLMModel.forward to ensure embedding outputs and all intermediate layer outputs are correctly returned
• Ensures proper handling and propagation of output_attentions, output_hidden_states, and return_dict flags

Why is this needed?

This aligns XLM with the standardized output tracing architecture introduced in #43979 ensuring consistent behavior across models and compatibility with the shared output capturing infrastructure.

[github-actions]

[For maintainers] Suggested jobs to run (before merge)

run-slow: flaubert, xlm