Add tutorial for custom checkpoints

Nobody has claimed this yet.

If I were to pretrain a model with my own data with something like Phi 1.5, how do I go about using the model? I have only a `last.ckpt` file containing all the weights and optimizers. The pretraining guide doesn't say how this file should be converted to something lit-gpt can use. I've also created the other config files, which the tutorial also lacks for a custom pretrained model. For instance, using my own tokenizer instead of the open source pretrained models'.

Start with the pretraining guide and the custom `last.ckpt` scenario described in the issue. Document how a checkpoint containing weights and optimizers can be converted for litGPT, and include the missing config-file and custom-tokenizer steps. Done means a user with their own pretrained model can follow the tutorial through loading and using it.

Written by the indexing model from the issue text.