Fix for multiblock GPU dists

[johnlees]

Rectangular parts were being transposed to the wrong size, overwriting some elements and leaving others as uninitialised

For future, to test this set an odd sample size and run with chunks = 3 in gpu_api.cpp

python sketchlib-runner.py query dist listeria.h5 --gpu 1

Other notes from build:

• Used gcc-11 from homebrew, but no longer necessary with newer cuda
• Need to be added to vglusers group to run CUDA now
• hdf5 LD_LIBRARY_PATH needs to be set to conda dir
• export SKETCHLIB_INSTALL=local, and set path and ld_library_path to include nvcc too

[nickjcroucher]

Code looks good, version bumped, I have installed and compiled, run it on a large dataset on GPU, and the output now behaves more sensibly, thanks for looking into this!