Support batch size scaling with dataloaders passed directly to `fit()`

🚀 Feature

This was first suggested/reported in #4000 and addressed in #4006 but then reverted in #4040 due to some GPU testing failure.

Currently, batch size scaling cannot be used with dataloaders passed directly to trainer.fit() as exception is raised at: https://github.com/PyTorchLightning/pytorch-lightning/blob/fe34bf2a653ebd50e6a3a00be829e3611f820c3c/pytorch_lightning/tuner/batch_size_scaling.py#L56-L58

The reason why this limitation was set is explained in #4006.

This limitation was set in place, to secure that the user had a field self.batch_size that we could alter. To remove this limitation we instead replaces the dataloader with a newly instantiated dataloader with altered batch size.

Motivation

Pitch

Alternatives

Additional context

• The limitation has been discovered in the current tuner revamp PR: https://github.com/PyTorchLightning/pytorch-lightning/pull/11089#discussion_r782849342.
• TODO after #11089 lands.

---

If you enjoy Lightning, check out our other projects! ⚡

• Metrics: Machine learning metrics for distributed, scalable PyTorch applications.

• Lite: enables pure PyTorch users to scale their existing code on any kind of device while retaining full control over their own loops and optimization logic.

• Flash: The fastest way to get a Lightning baseline! A collection of tasks for fast prototyping, baselining, fine-tuning, and solving problems with deep learning.

• Bolts: Pretrained SOTA Deep Learning models, callbacks, and more for research and production with PyTorch Lightning and PyTorch.

• Lightning Transformers: Flexible interface for high-performance research using SOTA Transformers leveraging Pytorch Lightning, Transformers, and Hydra.

cc @borda @justusschock @awaelchli @ninginthecloud @rohitgr7 @otaj @akihironitta