Fix labels & eos_token for SFT

[li-plus]

Fixed two issues:

• Padding should be ignored in training. Their labels should be set to -100 for CrossEntropyLoss to ignore them.
• Append correct eos_token to the response text. Otherwise the default <|endoftext|> will be tokenized into multiple tokens.

[li-plus]

Just rebased and fixed code format. Could anybody merge this PR since it's been approved. I don't have merge permission.

[tjruwase]

@li-plus, apologies for the delay on this. Can you please share a bit more details on the motivation for this PR. For example, what improvements did you observe in your workloads? Thanks!

[li-plus]

@tjruwase These are bugs that should be fixed. In master code, labels for padding are not ignored. Model will be trained to generate padding tokens, which is unnecessary. Also, eos token is not recognized as a single token. If you examine the input_ids in the training_scripts/opt/single_gpu/run_1.3b.sh example, it's tokenized into [49069, 15483, 1397, 1116, 29015, 15483, 15698], i.e. ['.<', '|', 'end', 'of', 'text', '|', '>'], which is wrong. This PR fixed these bugs.

[tjruwase]

@li-plus, got it. Thanks for this great contribution.