fix bug in run_clm_sft_with_peft.py

[STHSF]

Adjust pad token before count the number of tokens

[airaria]

We recommend using the Alpaca tokenizer when running run_clm_sft_with_peft.py.
The if statement is to check if the tokenizer is alpaca tokenizer (of which vocab size is 49954).

In #666, you used the merged tokenizer (of which vocab size is 49953), instead of the alpaca tokenizer from the chinese-alpaca-lora.
Therefore, if you switch to using the Alpaca tokenizer, the script should function correctly.