To better support OpenCL 2.0

Update OpenCL header files.

Fix FindOpenCL.cmake which doesn't find the OpenCL SDK on Windows

Resolves #260. Thanks to GitHub user @conradsnicta for reporting.

Problem was due to an overly aggressive clear() on the matrix_base.
Resolves #258.

Fix FindOpenCL.cmake which doesn't find the OpenCL SDK on Windows.  OPENCL_LIB_SEARCH_PATH is empty since it was never assigned and fools find_library.

allow for SVD and qr_method to be run on CPUs with stringent work size requirements. This corrects for invalid work sizes

Update documentation in ilu_operations.hpp

Resolves #257.

Clarified description.

add platform select to dense_blas

Thanks to sf.net user singam-sanjay for bringing this up:
https://sourceforge.net/p/viennacl/discussion/1143678/thread/d5f6a2fa/

CSR-data structure (row array) for ublas::compressed_matrix is not fully set up if some rows remain empty.
As a consequence, just copying over the row-array is not enough.
This commit replaces the plain copy with a manual build of the row array.

As a positive side effect, this will improve performance a (little) bit.

Remove the random_matrix/vector description in types.dox

Fix::a minimal example of Power Iteration does not work

The old behavior of the copy-CTOR was sharing data members rather than providing a clean copy.
This commit now provides a copy-CTOR, with the same the behavior with viennacl::matrix<T> (i.e. a clean copy, no data sharing).

Include build options into cache hash and avoid overwriting cache

Compile all kernels before caching program binary

This allows for GPU-accelerated application of ILUT.
The triangular solves are replaced by a truncated Neumann series with relaxation.
Enable via member function
 approximate_solves(iter)
of the viennacl::linalg::ilut_tag to run 'iter' relaxations.
Values in the range of 1-5 tend to give best results.

To better support the use case in #214.

Operations: element_prod(), element_div(), element_pow().
Only for float and double, no integer types supported (yet).

Operations: element_prod(), element_div(), element_pow().
Only for float and double, no integer types supported (yet).

The following new functions are now available:
 element_acosh()
 element_asinh()
 element_atanh()
 element_erf()
 element_erfc()
 element_exp2()
 element_exp10()
 element_log2()
 element_round()
 element_rsqrt()
 element_trunc()

Commit based on patch provided by Matthias Möller, authored by Andrzej Jaeschke <andrzej.jaeschke@gmail.com>
The original patch is extended to also clean up the CUDA backend -> less code :-)
C++11 guards are added to the tests.
Code compiles and tests run without C++11, but more testing desired to make sure everything is in order.

Reported-by: Breannan Smith via sourceforge.
             https://sourceforge.net/p/viennacl/discussion/1143678/thread/78f0a13b/