[Retiarii] support debug mode for easy debuggability

[QuanluZhang]

• support local debug mode. This mode only runs one trial locally, specifically, it applies mutators (default to choose the first candidate for the choices) to generate a new model, then run this model locally.
• make strategy the main thread for easily debugging user customized strategies
• update doc accordingly

[liuzhe-lz]

My major concern is that the feature is split into two modules, experiment and strategy, which increases the complexity and in my opinion, is unnecessary.
LocalDebugStrategy IS a strategy, so it seems intuitive to create a normal experiment with that strategy. But that's not the expected way to use it. Similarly, user may set a strategy before invoking experiment.local_debug_run(), and the strategy will be silently ignored.
Since current RetiariiExperiment API hides strategy.run() from end user, my suggestion is to define a "local debug experiment" and put all logic into that class or function. And in case we changed experiment API to let user invoke strategy.run() directly, I would suggest to make debug mode a static method.

[QuanluZhang]

My major concern is that the feature is split into two modules, experiment and strategy, which increases the complexity and in my opinion, is unnecessary.
LocalDebugStrategy IS a strategy, so it seems intuitive to create a normal experiment with that strategy. But that's not the expected way to use it. Similarly, user may set a strategy before invoking experiment.local_debug_run(), and the strategy will be silently ignored.
Since current RetiariiExperiment API hides strategy.run() from end user, my suggestion is to define a "local debug experiment" and put all logic into that class or function. And in case we changed experiment API to let user invoke strategy.run() directly, I would suggest to make debug mode a static method.

good points. To me, the current interface is the most convenient as it only requires changing one line of code. Maybe we have a follow-up pr for this.