fix: respect batch_size in v1 data_loader when permute=False

Apply changes from PR #391 by @MarcoGorworworelli:
- Fix train_gen() to iterate in proper batch_size chunks instead of
  yielding all time series at once when permute=False
- Add test_data_loader.py to verify batch boundaries

v1/src/timesfm/data_loader.py

@@ -149,11 +149,12 @@ class TimeSeriesdata(object):
     else:
       epoch_len = self.epoch_len
     for idx in perm[0:epoch_len]:
-      for _ in range(num_ts // self.batch_size + 1):
+      batch_indices = range(0, num_ts, self.batch_size)
+      for batch_idx in batch_indices:
         if self.permute:
           tsidx = np.random.choice(num_ts, size=self.batch_size, replace=False)
         else:
-          tsidx = np.arange(num_ts)
+          tsidx = np.arange(batch_idx, min(batch_idx + self.batch_size, num_ts))
         dtimes = np.arange(idx - hist_len, idx + self.pred_len)
         (
             bts_train,