Fix CRF label_weights: use register_buffer in __init__

Plain attribute assignment in __init__ conflicts with register_buffer()
in set_label_weights() — PyTorch raises KeyError because the attribute
already exists but isn't a buffer. Initialize as a None buffer instead.

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>

Author: Sid Mohan

pii-ner-v1/src/datafog_pii_ner/model/crf.py

@@ -21,7 +21,8 @@ def __init__(self, hidden_dim: int, num_labels: int, dropout: float = 0.1):
         self.num_labels = num_labels
         # label_weights: (num_labels,) tensor mapping label_id -> sequence weight.
         # Set via set_label_weights(). None means uniform weighting.
-        self.label_weights: torch.Tensor | None = None
+        # Use register_buffer so set_label_weights can update it later.
+        self.register_buffer("label_weights", None)
 
     def set_label_weights(self, weights: torch.Tensor):
         """Set per-label weights for tier-weighted sequence loss.