fix: improve row merging perf by 10x

The underlying GAPIC client uses protoplus for all requests and responses. However the underlying protos for ReadRowsResponse are never exposed to end users directly: the underlying chunks get merged into logic rows. The readability benefits provided by protoplus for ReadRows do not justify the costs. This change unwraps the protoplus messages and uses the raw protobuff message as input for row merging. This improves row merging performance by 10x. For 10k rows, each with 100 cells where each cell is 100 bytes and in groups of 100 rows per ReadRowsResponse, cProfile showed a 10x improvement:

old:          124266037 function calls in 68.208 seconds
new:          13042837 function calls in 7.787 seconds

google/cloud/bigtable/row_data.py

@@ -474,7 +474,11 @@ def _read_next(self):
 
     def _read_next_response(self):
         """Helper for :meth:`__iter__`."""
-        return self.retry(self._read_next, on_error=self._on_error)()
+        resp_protoplus = self.retry(self._read_next, on_error=self._on_error)()
+        # unwrap the underlying protobuf, there is a significant amount of
+        # overhead that protoplus imposes for very little gain. The protos
+        # are not user visible, so we just use the raw protos for merging.
+        return data_messages_v2_pb2.ReadRowsResponse.pb(resp_protoplus)
 
     def __iter__(self):
         """Consume the ``ReadRowsResponse`` s from the stream.
@@ -543,11 +547,12 @@ def _process_chunk(self, chunk):
     def _update_cell(self, chunk):
         if self._cell is None:
             qualifier = None
-            if "qualifier" in chunk:
-                qualifier = chunk.qualifier
+            if chunk.HasField("qualifier"):
+                qualifier = chunk.qualifier.value
+
             family = None
-            if "family_name" in chunk:
-                family = chunk.family_name
+            if chunk.HasField("family_name"):
+                family = chunk.family_name.value
 
             self._cell = PartialCellData(
                 chunk.row_key,
@@ -577,8 +582,8 @@ def _validate_chunk_reset_row(self, chunk):
 
         # No reset with other keys
         _raise_if(chunk.row_key)
-        _raise_if("family_name" in chunk)
-        _raise_if("qualifier" in chunk)
+        _raise_if(chunk.HasField("family_name"))
+        _raise_if(chunk.HasField("qualifier"))
         _raise_if(chunk.timestamp_micros)
         _raise_if(chunk.labels)
         _raise_if(chunk.value_size)