March backport changes

[mmdanziger]

This PR significantly simplifies the core modeling architecture by consolidating task-specific models (MLM, Sequence Classification, etc.) into a unified MultiTask architecture. It also introduces a user-friendly, Scanpy-style Python inference API, improves how model configurations are merged from checkpoints, and ensures backward compatibility for older checkpoints.

Key Changes

1. Unified MultiTask Architecture

• **Removed collation_strategy & ModelingStrategy**: The codebase no longer relies on tracking separate modeling strategies (mlm, sequence_classification, sequence_labeling). All models (SCBert, Performer, Nystromformer, Llama, ModernBert) now exclusively return and instantiate their ForMultiTaskModeling variants.
• HuggingFace AutoModel registrations have been updated so that both AutoModelForMaskedLM and AutoModelForSequenceClassification correctly resolve to the multitask models.
• Replaced dynamic training module resolution with direct usage of MultiTaskTrainingModule (or SequenceLabelingTrainingModule if enable_perturbation_metrics is set).

2. New Python Inference API

• Introduced bmfm_targets.inference.inference to allow seamless zero-shot predictions directly on AnnData objects.
• This feature acts similarly to a Scanpy tool (bmfm.inference(adata)), handling layer swapping, tokenization, prediction extraction, and appending the resulting embeddings and metadata labels back into the adata object automatically.

3. Robust Config & Checkpoint Merging

• Refactored SCBertMainConfig to handle config merging intelligently between YAML specifications and checkpoint hyperparameters (_merge_fields, _merge_label_columns, and _merge_configs_from_checkpoint).
• Clarified precedence: Checkpoint configs are largely authoritative during inference/prediction to match trained weights, while YAML configs take precedence/augment during training.
• Handles edge cases gracefully, such as clearing label_columns in predict mode if a checkpoint lacks label decoder weights.
• A generalized merge_configs function replaces merge_trainer_configs in task_utils.py.

4. Backward Compatibility & Migration

• Added migrate_checkpoint_if_needed to dynamically detect and convert legacy checkpoints (e.g., pure MLM or pure SeqCls) into the expected multitask format at runtime, eliminating the need to manually re-train old models.

5. Model & Layer Refinements

• Pooling Defaults: Changed the default pooling_method in TrainerConfig from "pooling_layer" to "first_token", which provides a safer default (especially for pure MLM models where the pooler is untrained).
• MVC Embeddings: Fixed missing dictionary unpacking for mvc_query_embeddings across multiple predictive layers, ensuring the correct query tensors are routed to MVC decoders.
• PEFT Support: Added output_attentions and return_dict arguments to forward signatures of model wrappers to preserve compatibility with standard PEFT/LoRA integrations.

6. Cleanup and Documentation

• Removed all references to collation_strategy across tutorial notebooks, GitHub Actions CI workflows, and .yaml config files.
• Updated README.md files to reflect the simplified data module configuration requirements.