Support an autotune policy between "none" and "quick"

[Chillee]

When designing a wrapper on top of autotuning, I want something that runs very quickly. However, using HELION_AUTOTUNE_EFFORT=minimal prevents kernel.autotune from working.

[oulgen]

@Chillee you can probably get this temporarily by changing defaults to fewer generations or larger ratio for convergence checking

for longer term, @choijon5 wanna take a look?

[choijon5]

You could limit the number of autotune generations with @helion.kernel(autotune_max_generations=N) or with env var HELION_AUTOTUNE_MAX_GENERATIONS=N. Otherwise, you could also directly set the constructor arguments as shown in: Direct Control Over Autotuners

Would these work for you @Chillee ?

[Chillee]

@choijon5 I think something a bit more convenient that folks could use by default would be useful here.

[choijon5]

I see. Would you like HELION_AUTOTUNE_EFFORT=minimal to do ~10% of full autotuning? quick does ~30%.

[Chillee]

@choijon5 I want it to do the minimal % of autotuning possible.