`lr_finder` fails when called after training for 1 or more epochs

Nobody has claimed this yet.

## 🐛 Bug

Calling [`lr_finder`](https://pytorch-lightning.readthedocs.io/en/latest/api/pytorch_lightning.tuner.lr_finder.html?highlight=lr_finder) on the model after `trainer.fit()` has been called will fail with:
```
LR finder stopped early due to diverging loss.
Failed to compute suggesting for `lr`. There might not be enough points.
```
, even when the default value of `min_lr=1e-08` has been changed to `1e-30`.

## Please reproduce using [the BoringModel and post here](https://colab.research.google.com/drive/1HvWVVTK8j2Nj52qU4Q4YCyzOm0_aLQF3?usp=sharing)

- Reproduced using a **callback**: https://colab.research.google.com/drive/1sbOPs8edyFi_idJNnd6gr3etyv7V57YU?usp=sharing
- Reproduced with **calling Trainer twice**: https://colab.research.google.com/drive/1WxUvayBBg_163nu8fjv-jsvPk6pUrSrK?usp=sharing

### To Reproduce

Add the following callback (as demonstrated with the BoringModel):

```
# Call Learning Rate finder after X epochs
class LRFinderXEpoch(Callback):
def __init__(self, epoch=1):
super().__init__()

self.epoch = epoch

def on_train_epoch_start(self, trainer, pl_module):
if trainer.current_epoch == self.epoch:
print("Calling learning rate finder!")
trainer.tune(pl_module)
# trainer.tuner.lr_find(pl_module, min_lr=1e-30)
```

### Expected behavior

Find the best learning rate after a few epochs of training (e.g. when doing Transfer Learning).

### Environment

```
* CUDA:
- GPU:
- Tesla T4
- available: True
- version: 10.1
* Packages:
- numpy: 1.18.5
- pyTorch_debug: True
- pyTorch_version: 1.7.0+cu101
- pytorch-lightning: 1.0.8
- tqdm: 4.41.1
* System:
- OS: Linux
- architecture:
- 64bit
-
- processor: x86_64
- python: 3.6.9
- version: #1 SMP Thu Jul 23 08:00:38 PDT 2020
```

### Additional context

Issue came from the following discussion: https://forums.pytorchlightning.ai/t/train-2-epochs-head-unfreeze-learning-rate-finder-continue-training-fit-one-cycle/366/4

Potentially related issues:
- https://github.com/PyTorchLightning/pytorch-lightning/issues/4784
- https://github.com/PyTorchLightning/pytorch-lightning/issues/4616

Start with the BoringModel reproducer and the LRFinderXEpoch callback, then compare trainer.tune(pl_module) with trainer.tuner.lr_find after Trainer.fit has run for one or more epochs. The change is complete when the learning-rate finder succeeds after resumed training, including with the documented callback scenario, and a regression test covers it.

Written by the indexing model from the issue text.