Update ipu_config.json

Set `replication_factor` and `inference_replication_factor` to dicts based on pod_type for flexibility. If no `pod_type` argument is provided, the default value of 4 will be retained.

ipu_config.json

@@ -7,7 +7,7 @@
   "execute_encoder_on_cpu_for_generation": false,
   "gradient_accumulation_steps": 2048,
   "inference_device_iterations": 5,
-  "inference_replication_factor": 4,
+  "inference_replication_factor": {"pod4": 1, "pod8": 2, "pod16": 4, "pod32": 8, "pod64": 16, "default": 4},
   "ipus_per_replica": 4,
   "layers_per_ipu": [
     0,
@@ -21,7 +21,7 @@
   "output_mode": "final",
   "recompute_checkpoint_every_layer": true,
   "replicated_tensor_sharding": true,
-  "replication_factor": 4,
+  "replication_factor": {"pod4": 1, "pod8": 2, "pod16": 4, "pod32": 8, "pod64": 16, "default": 4},
   "seed": 42,
   "sharded_execution_for_inference": false,
   "transformers_version": "4.20.1"