Call cudaFreeHost() on wrongPerGpu not cudaFree()

Update NCCL tests

Fix preprocessor version check for ncclGetLastError()

ncclGetLastError() was added in NCCL 2.13.0

Fix an issue with the last commit when data checking is disabled

Display N/A for error count in AlltoAll in-place test

AlltoAll does not support in-place buffers

Changed top-level Makefile behavior so that BUILDDIR is interpreted

as relative to top-level directory. This done is by abspath'ing it before
passing it to subdirectory Makefile's.

The old behavior had two cases: with and without BUILDDIR being set by
the user. With BUILDDIR not set, the build dir would be named "build"
in the top-level directory. If BUILDDIR was set, then the build dir
would be placed at "src/${BUILDDIR}".

The new behavior is simpler, if BUILDDIR is not set then it defaults
to "build", and the directory holding the final build is always at just
"${BUILDDIR}" in the top level.

Resync with NCCL 2.11.4

New operator: mulsum
New test: gather

Add CUDA graph support only for CUDA 11.3 and later builds

Fixes NVIDIA#90

Merge branch 'bfloat16'

Update PERFORMANCE.md