feat(models): multibackend all_to_all wrapper

[cathalobrien]

Small PR to fallback to support alltoall when using the Gloo backend to torch.distributed. This PR is needed to be able to run the transformer model on CPU. for 99.9% of users running on GPUs with the NCCL background, this change should not effect them

Gloo does not offer an alltoall primitive, as shown here

This PR implements am all_to_all fallback for Gloo, using the 'Linear Shift' algorithm from Hoffman and Rünger, 2013. Because of syntax for torch.dist changing in torch 2.6, older versions of torch are not supported.

[cathalobrien]

I written a script here to benchmark and test the correctness of this gloo alltoall with nccl.

python alltoall_test.py
Running on all-to-all on CPU over 4 members
Input list total size is 64.0MB
------------------------------  ------------  ------------  ------------  ------------  ------------  ------------
                          Name    Self CPU %      Self CPU   CPU total %     CPU total  CPU time avg    # of Calls
------------------------------  ------------  ------------  ------------  ------------  ------------  ------------
                    c10d::send         0.14%       1.722ms         0.14%       1.722ms     573.918us             3
                     gloo:send         0.00%       0.000us             0       13.540s        4.513s             3
                   c10d::recv_         0.00%      36.513us         0.00%      36.513us      12.171us             3
                     gloo:recv         0.00%       0.000us             0       18.687s        6.229s             3
            cudaGetDeviceCount         0.12%       1.441ms         0.12%       1.441ms     720.274us             2
------------------------------  ------------  ------------  ------------  ------------  ------------  ------------
Self CPU time total: 1.230s

Running on all-to-all on GPU over 4 members
Input list total size is 64.0MB
-------------------------------------------------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------
                                                   Name    Self CPU %      Self CPU   CPU total %     CPU total  CPU time avg     Self CUDA   Self CUDA %    CUDA total  CUDA time avg    # of Calls
-------------------------------------------------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------
                                        c10d::alltoall_         0.02%      71.892us        80.85%     272.351ms     272.351ms       0.000us         0.00%     129.453ms     129.453ms             1
                                     record_param_comms        49.68%     167.358ms        80.83%     272.284ms     136.142ms      64.727ms        99.45%     129.453ms      64.727ms             2
                                  cudaStreamIsCapturing         0.00%       3.088us         0.00%       3.088us       3.088us       0.000us         0.00%       0.000us       0.000us             1
                                        cudaEventRecord         0.00%      12.800us         0.00%      12.800us       2.133us       0.000us         0.00%       0.000us       0.000us             6
                                    cudaStreamWaitEvent         0.00%       9.823us         0.00%       9.823us       1.965us       0.000us         0.00%       0.000us       0.000us             5
-------------------------------------------------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------
Self CPU time total: 336.872ms
Self CUDA time total: 65.085ms

Does CPU and GPU output match? True

CPU runtime looks about an order of magnitude slower at modest input sizes, but the output is correct.