[Major] Add persistent scheduler for kernel management

[Enigmatisms]

Restructure the code into supporting multiple variants of tile schedulers (see renderer/scheduler.cuh). The original implementation is equivalent to the SingleTileScheduler. While I tried one persistent scheduler PreemptivePersistentTileScheduler, there is no speed bonus, since the single tile scheduler is already workload balanced. PreemptivePersistentTileScheduler works better in situations with many tiles with different workloads.

• This concludes the TODOs mentioned in [Feat] Improve the naive path tracer by implementing persistent kernels and schedulers #21.

Although the code changes are somewhat heavy, since there is no significant speed bonus, we will only publish a minor version release.

[Enigmatisms]

LGTM