[Feature] Flexible selection logic of execution plans

[dromniscience]

Thanks for this amazing project! When I use NCCL over MSCCL++ v0.6.0, I notice that the current implementation only considers the following condition for choosing an execution plan for a certain collective operation: bytes >= p.key.minMessageSize && bytes < p.key.maxMessageSize && inPlace == p.key.isInPlace. It does not consider whether NVLS is available, world size (since each execution plan only works for a particular world size), and many other factors. It would be better if there were an interface or some environment variables that allow users to control the execution plan.

[Binyang2014]

Thank you for the suggestion; we are actively updating the MSCCLPP DSL and execution-plan logic. We will consider providing interfaces to let user select the best execution plan. BTW, our new DSL design is on-going, you can checkout this branch feature/dsl and related PR #519 if you are interested.