Allow PaddingFree to work with DataCollatorForCompletionOnlyLM

[fabianlim]

Description

Currently padding free plugin only works with DataCollatorWithFlattening. This PR makes it also work with DataCollatorForCompletionOnlyLM that is used for the non-pretokenized use case

• in the scenarios.yaml, we should be above to use the chat_templates, but set tokenize=False and remove the null settings on the required fields for tokenization.
• verified that full-FT + paddingfree improvement is consistent at 22-23% for orca benches

Note:

• There are some inconsistent values in benchmarks for PF with FOAK tracked in Inconsistency in Padding-Free Benchmarks with Different Transformers Versions #70.

Tests on Flan Subset (6000 samples)

1. Verified that dataset data_cache.json is only formatted and untokenized. To ensure that the loss is masked, added a keyword 'RESPONSE:' in the chat template and as the response template that DataCollatorForCompletionOnlyLM will use to mask the loss.

   Example extracted from dataset['train']['output'][0]

   'Write the response. A 2 person conversation: -- Who was selected with the 5th pick in the 1974 NBA draft?. --  RESPONSE: Five other players from this draft, 6th pick Scott Wedman, 8th pick Campy Russell , 12th pick Brian Winters, 21st pick Billy Knight and 25th pick John Drew, were also selected to at least one All-Star Game.'
   

2. Verified that using an untokenized dataset to SFTTrainer matches previous padding-free performance with a pretokenized dataset.

   Untokenized FLAN Dataset

   Framework Config	Num Devices	Per Device Batch Size	Train Runtime (secs)	Speedups
   full-FT	2	4	1516	baseline
   padding-free	2	4	848	1.78x
   padding-free + multipack	2	4	747	2.02x

   Tokenized FLAN Dataset

   Framework Config	Num Devices	Per Device Batch Size	Train Runtime (secs)	Speedups
   full-FT	2	4	1537	baseline
   padding-free	2	4	859	1.79 x
   padding-free + multipack	2	4	751	2.05 x

[fabianlim]

@achew010 can you do a sanity check and open the data_cache.json to ensure it was not tokenized when tokenize=False