Rectify Missing Dataloader Preparation Call in PaddingFree Plugin Method

[achew010]

Description

This PR fixes a missing dataloader preparation call in the PaddingFree Plugin that will cause num optimization steps in distributed experiments to be computed wrongly.

With this fix, the relative performance improvement between padded vs padding-free approximately matches for

• FMS-Acceleration PaddingFree plugin, 24%
• Natively supported padding free in Transformers, 26%

ORCA Math 20K Subset - 1 Epoch

Transformers

Model	DataProcess	Grad Accum	Per Device Batch Size	Num Devices	Time	Throughput (token/s) / device
Mistral-7B	Padded	2	4	8	812	1216
Mistral-7B	Transformers Native PaddingFree	2	4	8	596	1658

FMS-Acceleration Padding-Free Plugin

Model	DataProcess	Grad Accum	Per Device Batch Size	Num Devices	Time	Throughput (token/s) / device
Mistral-7B	Padded	2	4	8	818	1208
Mistral-7B	PaddingFree Plugin	2	4	8	621	1568

[fabianlim]

Whats the difference between "Padded-Longest" and "Transformers Padded"? If there is no difference please edit the description to make it consistent