`trainer.fit_loop.setup_data()` does not refresh train dataset in `LightningModule`

Nobody has claimed this yet.

### Bug description

PR #16726 replaces the `reset_*_dataloader()` method calls with the respective `Loop.setup_data()` calls. This is also mentioned in the [migration guide](https://lightning.ai/docs/pytorch/stable/upgrade/from_1_9.html#developer).

However, on versions `<= 1.9`, calling `reset_train_dataloader()` would reinstantiate the dataloader from a `LightningModule`'s `train_dataloader()` method. This behaviour is now gone.

My specific use case is that I need to update the dataset of my model during training. I then use `on_train_epoch_end()` or a similar hook to call `reset_train_dataloader()`, to have the updated dataset in the next training epoch. I posted a minimal example below. You can run this example on both `v1.9` and `v2.0` to see the exact difference. `v1.9` runs without problems, whereas `v2.0` fails the second assertion in `training_step()`. I tested it on a fresh conda env install of both versions using python 3.10.

In case I am using the wrong `loop` to call `setup_data()` or am using the new interface incorrectly, please let me know. In that case I would also recommend providing some more hints in the migration guide or on PR #16726 since the current advice is not exactly clear. (i.e. which loops are "top level"?)

### What version are you seeing the problem on?

2.0+

### How to reproduce the bug

```python
try:
import lightning
except ModuleNotFoundError:
import pytorch_lightning as lightning

import torch
from torch.utils.data import DataLoader, TensorDataset

class Model(lightning.LightningModule):

def __init__(self):
super().__init__()
self.train_data = TensorDataset(torch.zeros(1, 1))

def configure_optimizers(self):
return None

def on_train_epoch_end(self):
self.train_data = TensorDataset(torch.ones(1, 1))

if int(lightning.__version__[0]) < 2:
# for version < 2.0 (works)
self.trainer.reset_train_dataloader()
else:
# for version >= 2.0 (does not work)
self.trainer.fit_loop.setup_data()

def train_dataloader(self):
return DataLoader(
self.train_data,
)

def training_step(self, batch, batch_idx):
# de-tuple
batch = batch[0]

if self.trainer.global_step == 0:
assert torch.allclose(batch, torch.zeros_like(batch))
else:
# this assertion fails on lightning v2.0
assert torch.allclose(batch, torch.ones_like(batch))

return torch.tensor(0.0, requires_grad=True)

model = Model()
trainer = lightning.Trainer(max_steps=2)
trainer.fit(model)
```

### Error messages and logs

```python
File "/home/lars/code/python/lightning-trainable/playground.py", line 38, in training_step
assert torch.allclose(batch, torch.ones_like(batch))
AssertionError
```

### Environment

Current environment

```
#- Lightning Component (e.g. Trainer, LightningModule, LightningApp, LightningWork, LightningFlow): Trainer, FitLoop, LightningModule
#- PyTorch Lightning Version (e.g., 1.5.0): 1.9 / 2.0
#- Lightning App Version (e.g., 0.5.2): -
#- PyTorch Version (e.g., 2.0): 1.9 / 2.0
#- Python version (e.g., 3.9): 3.10
#- OS (e.g., Linux): Ubuntu
#- CUDA/cuDNN version: 11.7
#- GPU models and configuration: RTX 2070
#- How you installed Lightning(`conda`, `pip`, source): conda
#- Running environment of LightningApp (e.g. local, cloud): local
```

### More info

_No response_

cc @borda @justusschock @awaelchli @carmocca

Start with PR #16726, the migration guide, and the Trainer's fit_loop.setup_data() entry point; compare it with the former reset_train_dataloader() behavior. Run the provided minimal example on Lightning 1.9 and 2.0, then verify that changing LightningModule.train_data refreshes the loader for the next epoch and that the supported usage is documented.

Written by the indexing model from the issue text.