[ROCm][Allreduce] Add dispatch mechanism for choosing performant allreduce implementations for AMD platforms

[zejunchen-zejun]

According to the different allreduce backends for ROCm devices, here a dispatcher mechanism has been developed for dispatching the performant allreduce implementations based on the different payload size, TP size and platforms. It has 3 main advantages:

• enhance the allreduce performance for workloads
• easy to maintain the different allreduce implementations and dispatch the performant one
• easily adaptable to new hardware(MI400 in future)

TODO:

• provide the performance data and accuracy with this PR on popular workloads

[mergify]

This pull request has merge conflicts that must be resolved before it can be
merged. Please rebase the PR, @zejunchen-zejun.

https://docs.github.com/en/pull-requests/collaborating-with-pull-requests/working-with-forks/syncing-a-fork

[gemini-code-assist]

Code Review

This pull request introduces a dispatch mechanism for selecting performant all-reduce implementations on ROCm platforms, which is a valuable performance enhancement. The overall structure is sound, but I've identified two critical issues that could lead to runtime errors due to improper handling of None values. Additionally, I've suggested adding missing type hints to improve code clarity and maintainability. Addressing these points will make the implementation more robust.