When using threads, one also has to link with chrono, date_time, and serialization.
Some of these components are not available in older Boost versions. What a mess.

Detected by Visual Studio 2010.

The first was due to the behavioral change of copy() in d81c7c57.
The second was a wrong work size, which was probably in there for a longer period of time already.

As a positive side-effect, this also improves multi-threaded operations by placing temporaries in the correct context.

Power iteration: Improved performance by avoiding a temporary, reduced tolerance for the tutorial for shorter execution times.

The power iteration implementation included the operation
 r <- prod(A, r);
which requires the creation of a temporary vector. Since this is called in a loop,
the overhead can become quite substantial, hence this was split into
 r2 <- prod(A, r);
 r  <- r2;

Did not consider matrix padding.

The lengths of the temporary buffers wasn't computed correctly, did not consider the new matrix padding.

Exceptions: Now throwing cuda_not_available_exception() when trying to use CUDA functionality without VIENNACL_WITH_CUDA.

This tutorials hasn't been modified to consider the matrix padding.

Newer versions of Boost require to link with libboost_serialization when using uBLAS.
While this was automatically resolved on Linux, it is not the case with Visual Studio.
Linkage is now added, so tests should compile nicely now.

The use of 'uint' in CUDA kernels does not work with Visual Studio 2010 (and probably other versions as well),
hence this change.

Now a copy() of generic->ViennaCL-type resizes the ViennaCL object if it has zero size.
Conversely, a copy() from ViennaCL-type->generic checks for the correct size of the generic object,
but no longer resizes anything. The rationale is that ViennaCL cannot reasonably know how to resize() external objects of arbitrary type.
Hence, the user is supposed to resize the type, because (s)he easily knows how to do this.

The only warning left is a warning regarding unsigned integer sign flips, which will be addressed after 1.5.0 is out.

Apparently hasn't been enabled for quite a while. Seems like it was simply forgotten.

The repetitive bit-shifting has now been moved to a single implementation for CUDA and one for OpenCL.
Also, minor CUDA warnings fixed along the way.

Visual Studio does not export functions by default, but rather require
  __declspec(dllexport)
and
  __declspec(dllimport)
to be set accordingly. These export declarations now get set when using CMake.

A reference to a temporary was returned, which is likely to cause troubles.

It seemed like Doxygen had some troubles with some symbol shortcuts
obtained from 'using namespace ...' in the generator implementation.
Hence, I removed these using-declarations and added full namespaces to the respective type names.

Two unused member variables and one potential use of an uninitialized variable.