fix: Resolve race condition in concurrent Transaction initialization

Implement a double-checked locking pattern in Transaction and _SnapshotBase methods.
When multiple threads attempt to use a lazy transaction simultaneously, they race to acquire the lock.
Previously, losing threads would acquire the lock and blindly send another 'begin transaction' request,
ignoring that the winner had already initialized the transaction ID.

This change ensures that threads re-check self._transaction_id after acquiring the lock.
If the ID is present, they skip the initialization request and use the established ID.

Author: aseering

google/cloud/spanner_v1/snapshot.py

@@ -638,6 +638,10 @@ def _get_streamed_result_set(
         if self._transaction_id is None:
             is_inline_begin = True
             self._lock.acquire()
+            if self._transaction_id is not None:
+                is_inline_begin = False
+                self._lock.release()
+                request.transaction = TransactionSelector(id=self._transaction_id)
 
         iterator = _restart_on_unavailable(
             method=method,

google/cloud/spanner_v1/transaction.py

@@ -512,6 +512,9 @@ def execute_update(
         if self._transaction_id is None:
             is_inline_begin = True
             self._lock.acquire()
+            if self._transaction_id is not None:
+                is_inline_begin = False
+                self._lock.release()
 
         execute_sql_request = ExecuteSqlRequest(
             session=session.name,
@@ -670,6 +673,9 @@ def batch_update(
         if self._transaction_id is None:
             is_inline_begin = True
             self._lock.acquire()
+            if self._transaction_id is not None:
+                is_inline_begin = False
+                self._lock.release()
 
         execute_batch_dml_request = ExecuteBatchDmlRequest(
             session=session.name,