Automatic logger removal if the training failed (error happening on the sanity check)

Nobody has claimed this yet.

## 🚀 Feature
Let lightning automatically remove Tensor board logger folder and files in case the training wasn't started successfully due to an error.

### Motivation
Right now lightning creates a logger file and then runs sanity checks ect.
If there are any bugs in the code, every run creates a new event file and a folder for it.
This leads to a stack of empty logger files that pile up and are cumbersome to remove manually, especially on a remote machine.

### Pitch
Logger folder gets removed if the error occurs before first successful training step in the epoch :)

cc @borda @awaelchli @Blaizzy

The issue names no files, tests, or entry points. Start by locating TensorBoard logger creation and the sanity-check/training lifecycle; done means a failed pre-training run removes its logger folder and files while a run that reaches a successful training step preserves them.

Written by the indexing model from the issue text.