README: Updated version for GitHub to reflect that there is no more OpenCL kernel file generation necessary.

The rationale is that developers want to include the tests in their builds to make sure no API functionality is broken.
Similarly, package maintainers want to run their tests against the source release.
Only for the 'standard release' one cannot assume that Boost is available, hence tests should be switched off by default.

Now:
 C = prod(A, B); C = prod(A, trans(B));
fully supported, where A is sparse and B, C are dense (both row- or column-major).
Some kernels can be further tuned towards improved coalesced memory transfers,
or for avoiding memory transfers at all.

CUDA needs to be added later.

SpMdM implementation for CSR, COO, and ELL now complete.

CPU and OpenCL backends pass tests, CUDA implementation still needed.

CPU version working, OpenCL and CUDA versions work for row-major only.

CPU version was already prepared. Added OpenCL and CUDA implementation.
This commit fixes also fixes recent problems of coordinate_matrix-times-vector for CUDA.

Instead of a plain 'int', the ViennaCLInt typedef is now used.

The code for this was fixed/added already here:
deae066e

Routines are not supposed to dump status messages to terminal.

The use of random numbers resulted in some unpredictable failures in the test because of numerical instabilities.
This is now fixed by adding extra weight to the matrix diagonal and reducing the range of random values.

Older versions of Boost (e.g. 1.33) are only able to call a functor,
hence changed tutorial to a functor rather than a separate function.

On certain older systems (e.g. CentOS 5.x) there is no CMake 2.8.x available,
so we shouldn't make things more complicated than necessary.

The tests require uBLAS to be available, which is not necessary the case on the target machine.
If tests should be build, configure with
 cmake .. -DENABLE_UBLAS=ON -DBUILD_TESTING=ON

Filesystem is no longer required, thus removed.
System was introduced with 1.35.0, so the link would file with older versions of Boost.

In some situation an entry might have been computed already, but then be cleared instead of the other way round.
This is now fixed, tests no longer fail once in a while.

For matrices with less than 64 rows, the group boundaries were calculated incorrectly, thus producing wrong results.
Moreover, in some corner cases the last entry computed by a group may not have been written to the result vector.
Both issues are supposed to be corrected, tests pass with a recent Intel OpenCL SDK on a CPU.
Tests on GPUs still required.

This should allow us to conveniently switch to signed integer types if needed.

The permutation array returned by viennacl::reorder() routines now maps the old node index to the new node index,
i.e. permutation[old_index] = new_index.
This is more appropriate for use in other algorithms, where one had to frequently invert the permutation so far.

The performance gain was mainly obtained through replacing std::vector<> by std::pair<> in a container.
In addition, integer conversions are now removed, resulting in less conversion warnings at -Weverything at Clang.

libviennacl: Introduced ViennaCLInt as main integer type for OpenMP and OpenCL, not using any size_t anymore

Since Fortran has issues with unsigned integers, size_t will sooner or later cause troubles.
For flexibility, the integer type is now typedef'd to ViennaCLInt, so one can quickly change the
main integer type if needed.
The choice of 'int' was motivated by CUBLAS, where functions also take 'int' as integer parameter type.

fix a few implicit conversion warnings

Although they are allowed as optional by the standard in some circumstances, some older compilers and static code analyzers may complain.

Still transposition of the block factors required, which is way too slow at the moment.

The function was apparently implemented, but accidentally not included in the program string.

Two exceptions:
 - viennacl/generator: map_functor.hpp leads to a circular dependency, so cannot (and should not) be included directly.
 - viennacl/rand:      Currently not in use, prepared for inclusion in a later release.

Fixes #48