(Fix)(app) Delay Imports that require torch for app launch

[mickr777]

Summary

Fix for RuntimeError: configure_torch_cuda_allocator() must be called before importing torch. on app start

Related Issues / Discussions

Traceback (most recent call last):
  File "D:\InvokeMain\InvokeAI\scripts\invokeai-web.py", line 20, in <module>
    main()
  File "D:\InvokeMain\InvokeAI\scripts\invokeai-web.py", line 16, in main
    run_app()
  File "D:\InvokeMain\InvokeAI\invokeai\app\run_app.py", line 33, in run_app
    configure_torch_cuda_allocator(app_config.pytorch_cuda_alloc_conf, logger)
  File "D:\InvokeMain\InvokeAI\invokeai\app\util\torch_cuda_allocator.py", line 13, in configure_torch_cuda_allocator
    raise RuntimeError("configure_torch_cuda_allocator() must be called before importing torch.")
RuntimeError: configure_torch_cuda_allocator() must be called before importing torch.

• The PR has a short but descriptive title, suitable for a changelog

[psychedelicious]

I can't reproduce this locally - maybe it's platform specific or related to certain custom nodes? At any rate, thanks for the fix!