Introducing coarse-grained molecular dynamics simulations and updates to training

[datagato]

High-level updates from this PR:

• Introduce coarse grained MD simulations using the Martini 3 force field (with solvent subtraction) for faster Proteogram generation (5-10x faster than atomistic)
• Fixes to atomistic Proteogram creation for more consistent channel logic and stacking
• New training script that does train/test split on-the-fly with seed argument for reproducibility (this can be used for k-fold cross-validation later on by simply changing the seed); also introduces use of focal loss (a flag on the command line) for the training loss function
• PDB download script by date cutoff for future data pulls (when going beyond SCOPe)
• Annotation script updates (want a "GO slim" as a future label for Proteograms for classification according to the Gene Ontology annotations (biological process, molecular function and cellular component)
• Triplet-hierarchy metric learning (new proteogram/v2/losses.py): HierarchicalTripletLoss
  with SCOPe-distance-scaled margins + HierarchicalPKSampler (optional fold/superfamily
  stratification). Wired into the train script as --loss triplet_hierarchy, producing an
  L2-normalized embedding model trained and early-stopped on validation Precision@K rather
  than a classifier. Fully parameterized (embed_dim, triplet_p/k, margins, base_margin,
  families_per_stratum, min_family_size, stratify_level, early_stop_level) with collapse
  diagnostics (pos/neg distance, valid_anchor_frac, paired_frac). This is an optional path and the previous usage still applies.
• ViT-B/16 architecture option (--model vit) alongside the from-scratch CNN and ResNet18,
  including fixed 224x224 resize handling (224x224 is required for this model so all proteograms must be resized) and trailing-block fine-tuning.
• Self-describing checkpoints + padding consistency: Img2Vec auto-detects the {state_dict,
  meta} embedding format and exposes embedding_meta; measure_similarity_v2 and
  query_similar_proteins now derive the search-time pad target (and query length cutoff) from
  the checkpoint's recorded max_image_size, falling back to a new config search_pad_size.
  Fixes a silent train/eval mismatch that cropped/rescaled larger proteins at search time.
• Bug fixes: DataLoader crash when --batch_size is omitted (config fallback wasn't written
  back to args, yielding batch_size=None), and roc_auc_score failure on 2-class eval sets.
• Documentation updates