Fix THROUGHPUT_START_STEP

Author: mathemakitten

tests/functional_tests/test_cases/gpt/gpt_grpo_tp4_pp1_dp2_8b_cudagraphs_throughput/model_config.yaml

@@ -1,5 +1,6 @@
 ENV_VARS:
   CUDA_DEVICE_MAX_CONNECTIONS: 1
+  THROUGHPUT_START_STEP: 10
   NVTE_ALLOW_NONDETERMINISTIC_ALGO: 0
   NCCL_ALGO: Ring
   CUBLAS_WORKSPACE_CONFIG: :4096:8
@@ -100,5 +101,3 @@ METRICS:
   - "mem-allocated-bytes"
   - "mem-max-allocated-bytes"
   - "iteration-time"
-THROUGHPUT_TEST_PARAMS:
-  --start_step: 10