fix: handle control pool smaller than n_samples in _sample_consecutive_controls

When n_samples > pool_size (e.g., observational data with rare cell types
having only 2 control cells but sentence_len=64), the old tail+head wrap
only wrapped once, returning fewer elements than requested. This caused
IndexError in __getitems__ during multi-worker DataLoader training.

Use modular arithmetic to wrap around the pool as many times as needed.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

Author: abhinadduri

src/cell_load/mapping_strategies/random.py

@@ -276,6 +276,7 @@ def _sample_consecutive_controls(
         if start + n_samples <= pool_size:
             return np.array(pool[start : start + n_samples], dtype=np.int64)
 
-        tail = pool[start:]
-        head = pool[: n_samples - len(tail)]
-        return np.array(tail + head, dtype=np.int64)
+        # Wrap around the pool as many times as needed
+        pool_arr = np.array(pool, dtype=np.int64)
+        indices = np.arange(start, start + n_samples) % pool_size
+        return pool_arr[indices]