multi-GPU support v2

[winglet0996]

Hi! This PR is a tested multi-GPU inference practice with a test script test_multigpu.py (not a standalone unit test yet and need to run locally with a folder containing WSIs).
Our test result on 6 TCGA WSIs is as follows:

SCENARIO	TIME(s)
1GPU_NoCache	74.07
1GPU_Cache	66.42
2GPU_NoCache	48.30
2GPU_Cache	47.30

Features:

1. Enabling both WSI cache and non-cache mode with --gpus argument, by distributing all slide tasks equally on all allocated GPUs.
2. Auto skip completed slides when assigning tasks.
3. Clean stale .lock files and WSI caches before a run.
4. Support loading local Gigapath slide_encoder from local local_ckpts.json.

Feel free to test on your side to see if any bugs.
Thx!