update default resnet LOSS_SCALER to 256 [pr] (tinygrad#8774)

Author: chenyuxyz

examples/mlperf/model_train.py

@@ -79,7 +79,7 @@ def train_resnet():
   lr_warmup_epochs  = config["lr_warmup_epochs"]  = getenv("WARMUP_EPOCHS", 2)
   decay             = config["decay"]             = getenv("DECAY", 2e-4)
 
-  loss_scaler       = config["LOSS_SCALER"]       = getenv("LOSS_SCALER", 128.0 if dtypes.default_float == dtypes.float16 else 1.0)
+  loss_scaler       = config["LOSS_SCALER"]       = getenv("LOSS_SCALER", 256.0 if dtypes.default_float == dtypes.float16 else 1.0)
 
   target, achieved  = getenv("TARGET", 0.759), False
   eval_start_epoch  = getenv("EVAL_START_EPOCH", 0)