Enable TensorPipe CUDA sending to self

[mrshenli]

Stack from ghstack:

• Let TensorPipe detect peer access #50676 Let TensorPipe detect peer access
• Enable TensorPipe CUDA fallback channel #50675 Enable TensorPipe CUDA fallback channel
• Enable TensorPipe CUDA sending to self #50674 Enable TensorPipe CUDA sending to self

Differential Revision: D25941964

[codecov]

Codecov Report

Merging #50674 (ed9ede9) into gh/mrshenli/284/base (8b501df) will increase coverage by 56.15%.
The diff coverage is 72.72%.

@@                    Coverage Diff                    @@
##           gh/mrshenli/284/base   #50674       +/-   ##
=========================================================
+ Coverage                 24.49%   80.65%   +56.15%     
=========================================================
  Files                       487     1912     +1425     
  Lines                     61361   208111   +146750     
=========================================================
+ Hits                      15033   167860   +152827     
+ Misses                    46328    40251     -6077     

[facebook-github-bot]

@mrshenli merged this pull request in 94d9a7e.