Load transformer models saved on GPU on CPU-only machines

[maxteneff]

Description

Fixes #286.

A fitted SASRecModel or BERT4RecModel saved on a GPU machine can't be loaded with load_model, load or loads on a machine without CUDA:

RuntimeError: Attempting to deserialize object on a CUDA device but torch.cuda.is_available() is False. If you are running on a CPU-only machine, please use torch.load with map_location=torch.device('cpu') to map your storages to the CPU.

A pickled fitted model holds its Lightning checkpoint, which TransformerModelBase.__setstate__ reads with torch.load(...) without a map_location. torch then tries to restore the tensors on the device they were saved on. Unlike load_from_checkpoint, load gives users no way to pass a map_location.

This PR makes __setstate__ load the checkpoint to CPU. Nothing changes on GPU machines: TransformerLightningModule._prepare_for_inference moves the model to recommend_torch_device (CUDA when it's available) before every inference, and the trainer places the model on its device when fitting.

Type of change

• Bug fix (non-breaking change which fixes an issue)
• New feature (non-breaking change which adds functionality)
• Breaking change (fix or feature that would cause existing functionality to not work as expected)
• Optimization

How Has This Been Tested?

• Added test_load_fitted_model_saved_on_gpu_without_gpu for SASRecModel and BERT4RecModel. It saves a fitted model with tensors tagged as cuda:0 (patching torch.serialization.location_tag), makes torch.cuda.is_available() return False, loads the model, and checks that its recommendations match the original model. Without the fix it fails with the error above; with the fix it passes.
• tests/models/nn/transformers: 213 passed, 57 skipped (GPU-only).
• isort, black, flake8, pylint, mypy and codespell pass on the changed files.
• I don't have a CUDA machine, so I haven't tried a checkpoint saved on a real GPU. The test reproduces the same code path by tagging the tensors as CUDA.

Checklist:

• Read the contribution guide: yes
• Docstrings: no public API changes
• New tests: added
• Changelog: updated

Not included here: load_weights_from_checkpoint also calls torch.load without a map_location, so a GPU checkpoint passed to it fails the same way on a CPU-only machine. Happy to handle it in a follow-up, e.g. by adding a map_location argument like load_from_checkpoint has.