4-IPU base config

Uses less IPUs than current model-hub config file. This does use a different assignment method in parallelisation, which will be PRd separately.

ipu_config.json

@@ -1,16 +1,16 @@
 {
-    "optimizer_state_offchip": true,
+    "optimizer_state_offchip": false,
     "replicated_tensor_sharding": true,
     "enable_half_first_order_momentum": true,
     "enable_half_partials": true,
     "recompute_checkpoint_every_layer": false,
-    "device_iterations": 2,
+    "device_iterations": 10,
     "inference_device_iterations": 4,
     "replication_factor": 2,
     "inference_replication_factor": 2,
-    "gradient_accumulation_steps": 16,
+    "gradient_accumulation_steps": 32,
     "executable_cache_dir": "./exe_cache",
-    "ipus_per_replica": 8,
-    "layers_per_ipu": [2, 1, 4, 3, 3, 3, 3, 0],
-    "matmul_proportion": [0.1, 0.1, 0.1, 0.3, 0.3, 0.3, 0.3, 0.1]
+    "ipus_per_replica": 4,
+    "layers_per_ipu": [6, 5, 5, 7],
+    "matmul_proportion": [0.2, 1.0, 1.0, 1.0]
 }