Various optimizations & fix YOLOv8 image value scaling bug

[seth-planet]

• Increase parallelism count as tpu_runner should be thread-safe. Threads are cheap so we can have a large value here. (?)
• Fix minor bugs when running TPU code in isolation.
• Delete unused file.
• Reduce mem copies & GIL lock in TPU running loop.
• Convert output to float32 during output tensor copy to reduce memory churn.
• Rework watchdog timing & handling due to complaint about TPU re-init time.
• Convert input image to RGB only if needed.
• Revert input pixel value scaling optimization that broke YOLOv8.
• Input image auto-contrast & optimizations.