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.

Jakub Klinkovský authored Aug 11, 2019 and Jakub Klinkovský committed Aug 17, 2019

They are not implemented anyway...

This is necessary to be able to run tests in parallel.

The build takes too long because of this and nobody uses it anyway. CUDA
does not support long double in device code at all.

NDArray

See merge request !18

DistributedNDArray: added methods getLocalIndexer, getLocalView, getConstLocalView, getStorageIndex and getData

This is necessary for local indexing of DistributedNDArray.