It caused a weird compiler error when compiled with g++ -Werror.

And running CUDA binaries doesn't crash anymore, though they still do
not produce any meaningful coverage report, even for host code. This is
due to nvcc's temporary source files, we will have to compile with
clang++ natively to fix this.

They lead to internal compiler error in some cases.

Documentation: fixed branch name for CI builds

See merge request !40

CI: deploy with artifacts

See merge request !39

Setup for automatic documentation deployment

See merge request !38

Reduction and multireduction refactoring

Brief summary:

- rewritten multireduction using lambda functions
- avoided `volatile` using `__syncwarp()`
- using reduction functions with `return a + b` instead of `a += b`
- using `std::plus`, `std::multiplies`, `std::logical_and`, `std::logical_or`, etc. instead of custom lambda functions
- optimized OpenMP thread counts for reduction and multireduction
- added computation of sample standard deviation to benchmarks
- implemented parallel prefix-sum with OpenMP
- implemented distributed prefix-sum

See merge request !37

Fixes #43

PrefixSum: separate first and second phase for OpenMP implementation and expose performFirstPhase and performSecondPhase methods

Same changes as for the regular Reduction operation...

Fixes #42

Benchmarks can be easily profiled even without this parameter, so it was
just an unnecessary complication.

This is nicer because it more clearly separates data load, computation
and data store. Furthermore, it allows to use instances of std::plus,
std::logical_and, std::logical_or, etc. instead of custom lambda
functions.

It contained only methods for prefixSum and segmentedPrefixSum, which
were identical for Host and Cuda, so they can be easily implemented
directly in Vector and VectorView.

The performance seems to be identical to the code using volatile.