[Auto Parallel] Support semi-auto trainer and fit Llama2 training (#7885)

* support semi-auto trainer and fit Llama2 training

* support shard_dataloader in dynamic semi-auto

* rewrite traning loop

* refactor traning loop

* refine args of auto trainer

* broadcast loss

* add auto ci cases

Author: haohongxiang

llm/llama/auto_parallel/run_auto.sh

@@ -68,6 +68,6 @@ python -u -m paddle.distributed.launch \
     --do_eval \
     --device "gpu" \
     --data_impl "mmap" \
-    --parallel_mode "auto"
+    --enable_auto_parallel 1
 
     # --resume_from_checkpoint "output/llama_auto_serial/checkpoint-2" \

llm/llama/auto_parallel/run_auto_sp.sh

@@ -68,7 +68,7 @@ python -u -m paddle.distributed.launch \
     --do_eval \
     --device "gpu" \
     --data_impl "mmap" \
-    --parallel_mode "auto" \
+    --enable_auto_parallel 1 \
     --sequence_parallel true \
 
     # --resume_from_checkpoint "output/llama_auto_serial/checkpoint-2" \