Fix the hash_size_cumsum cond in TBE's init (pytorch#1862)

Summary:
Pull Request resolved: pytorch#1862

This diff fixes `ValueError: math domain error` introduced by
D46618714

Reviewed By: q10, jianyuh, yuguo68

Differential Revision: D47236684

fbshipit-source-id: d5375928ad5ee6bad58c183355b52fac71b4b211

fbgemm_gpu/fbgemm_gpu/split_table_batched_embeddings_ops_training.py

@@ -432,7 +432,7 @@ def __init__(  # noqa C901
         )
         hash_size_cumsum = [0] + list(accumulate(rows))
         self.total_hash_size: int = int(hash_size_cumsum[-1])
-        if hash_size_cumsum == 0:
+        if self.total_hash_size == 0:
             self.total_hash_size_bits: int = 0
         else:
             self.total_hash_size_bits: int = int(log2(float(self.total_hash_size)) + 1)