Phase 3 - Always On Tensorboard fixes.

[O-J1]

Pasted from my notes when I had to split the PRs:

• Tensorboard should be controlled from a single dropdown
• Tensorboard creation should be debounced.
• If Tensorboard is set to training, it must terminate when training finishes.
• If TB mode is always-on it must not restart once training finishes, it must persist
• Tensorboard locally and Cloud TCP tunnelling should still work.
• Cloud Training may not cause file/path doesnt exist warnings, only invalid characters