Revert "use send/recv imple all2all (#1351)"

[Chao1Han]

This reverts commit 7ee006f.

[Copilot]

Pull Request Overview

This PR reverts a previous commit that replaced the send/recv alltoall implementation with an alltoallv-based approach while introducing a helper function for flattening tensors.

• Reintroduces a different alltoall implementation using ccl::alltoallv.
• Adds a helper function (computeLengthsAndCheckAndGetFlat) to compute flattened tensor lengths and determine if tensors are already contiguous.

Comments suppressed due to low confidence (2)

src/xccl/ProcessGroupXCCL.cpp:43

• Consider adding a check to ensure that the 'tensors' vector is not empty before accessing its first element, to avoid potential out-of-range errors.

bool computeLengthsAndCheckAndGetFlat(const std::vector<at::Tensor>& tensors, std::vector<size_t>& lengths, at::Tensor& flatTensor, int64_t& flatLength) {

src/xccl/ProcessGroupXCCL.cpp:2000

• Using a reinterpret cast from 'size_t*' to 'int64_t*' might introduce issues on platforms where these types differ; consider using a conversion that safely translates the vector elements to int64_t.

auto flatInputSplits = flatInput.split_with_sizes(c10::IntArrayRef((int64_t*)sendCounts.data(), sendCounts.size()), 0);