Allreduce implementation, various smaller extra changes

[keshprad]

Summary of changes:

• New collective for AllReduce: implemented in both Python (pccl/all_reduce.py) and C++ (csrc/all_reduce.cpp/h).
  • Recursive and ring algorithms available through a unified _all_reduce / all_reduce_2D Python interface and a all_reduce_mpi C++ pybind11 binding.
• Added missing: Ring AllGather
• int → int64_t type fixes: Updates element count and block size parameters across all_gather, reduce_scatter, and vector add kernel to prevent overflow when operating on large tensors.