[RLlib] Change the slateq regression learning test with GPU to use torch only (#23168)

release/rllib_tests/learning_tests/hard_learning_tests.yaml

@@ -468,6 +468,11 @@ slateq-interest-evolution-recsim-env:
         env_config:
             # Env class specified above takes one `config` arg in its c'tor:
             config:
+                # GPU support currently does not work for tensorflow. For the learning
+                # test for now, we'll use pytorch, and investigate why tensorflow is not working.
+                # One indication of the issue is that the loss inputs are not flattened from the
+                # original dictionary spaces.
+                framework: torch
                 # Each step, sample `num_candidates` documents using the env-internal
                 # document sampler model (a logic that creates n documents to select
                 # the slate from).