add rm datadir before multiple datasets training to avoid repeating samples

convlab2/base_models/bert/train_bio.sh

...ase_models/gpt/keyword_extraction/gen_pretraining_data.sh

convlab2/base_models/t5/dst/run_dst_pretrain.sh

convlab2/base_models/t5/nlg/run_nlg_pretrain.sh

convlab2/base_models/t5/nlu/run_nlu_pretrain.sh

convlab2/base_models/t5/rg/run_rg.sh