Make LightningIR compatible with transformers v5

[RaykKretzschmar]

This PR updates Lightning IR for compatibility with Transformers v5.

The main changes are:

• Update dynamic config/model/tokenizer class factories for Transformers v5 behavior.
• Preserve and resolve backbone_model_type more robustly when loading derived Lightning IR configs.
• Adapt model-specific logic for v5, including SPLADE tied-weight metadata and SetEncoder attention mask handling.
• Update tokenizer handling for newer tokenizer mappings and T5-style cross-encoder formatting.
• Normalize dataset training targets to float32.
• Add compatibility shims for external model comparison tests where upstream libraries still assume older Transformers internals.
• Stabilize shared test configuration and update typing for the Python 3.10+ baseline.