Refactor SF packing

1) Separate out local communication from remote communication
2) Directly pass root/leafdata to MPI when suitable

Merge branch 'stefanozampini/sfwin-reuse-win' into 'master'

Improve PETSCSFWINDOW

See merge request petsc/petsc!2336

Merge remote-tracking branch 'origin/maint'

Fix 'may be used uninitialized' compiler warnings

PETSCSFBASIC: only reset BASIC specific data when destroying the object

PetscSFBcastAndOpBegin_Allgatherv: fix gcc-8 'may be used uninitialized'

Merge branch 'stefanozampini/petscpartitioner-target-weights' into 'master'

Fix PetscPartitioner and add support for target partition weights

See merge request petsc/petsc!2333

PetscSortedIntRemoveDups: use PetscCheckSorted

Merge branch 'jczhang/feature-sf-do-pack-on-gpu' into 'master'

Add support to do pack/unpack on GPU and do MPI on CPU

See merge request petsc/petsc!2205

Merge branch 'maint'

PetscSF: Fix buffer displacements for 32bit integer builds

Use MPI_Aint to store displacements and prevent truncation.
This issue occurs for buffer sizes in bytes approaching the 2GB limit,
`bufsize

PetscSF: Fix buffer displacements for 32bit integer builds

Use MPI_Aint to store displacements and prevent truncation.
This issue occurs for buffer sizes in bytes approaching the 2GB limit,
`bufsize=n*unitbytes` with `n<=INT_MAX` but `n*unitbytes>INT_MAX`.

show more ...

Merge branch 'maint'

Fix a bug: selfbuflen is in <unit>

Add an option -sf_use_pinned_buffer to use non-pagable host memory for send/recv buffer when passing GPU data

Log transfer between CPU and GPU

Fix a performance bug

Clarify a comment

Rename rkey, lkey to rootdata, leafdata

Support data on device but no gpu-aware MPI for sf(all)gather(v)

Fix a typo: selfbuflen should be in terms of <unit>

Renamed PetscSFPackWaitall_Basic to PetscSFPackWaitall

Support "pack on GPU, do MPI on CPU" for SF Neighbor

Add option that lets users do packing on GPU and do MPI on CPU

Merge branch 'maint'

Spelling: Destory -> Destroy

Commit-type: style-fix