Fix reduce-by-key CUDA (#737)

Fix bug in CUDA reduce_by_key implementation on V100 or later GPUs.
The bug was not triggered for current use cases of the algorithm in Arbor, though it will be a problem when more than one reduction is to be performed in a single kernel invocation, which is required for ac
cumulating both current and conductance values.
* Use warp-synchronous aware operations to avoid problems on V100.
* Simplify reduction kernel.
* Rename `run_length` ancillary data structure to `key_set_pos`.
* Add unit tests that trigger the incorrect behaviour observed in #736.

Fixes #736.