lcov takes much too long making the travis builds time out and fail.

Transposed forward solves accessed invalid shared memory, potentially
causing runtime failures on some GPUs. This is now fixed,
thanks to oclgrind for locating the problem quickly.
Addresses the failures Philippe mentioned in #105.

Definitely needs better testing, this was only caught in triangular solvers.

Reset local work size to 1 as CPU default.
Moved a barrier outside an if-conditional in order to simplify the job for the compiler.

local work sizes 1 or 2 for CPUs seem to cause problems with some SDKs.

Parts of the test consists of only taking timings.
No need to spend a lot of time on these for nightly tests,
better integrate into a separate benchmark suite.

Running with local work size 1 apparently leads to failures on
both AMD and Intel SDKs (maybe due to barriers being ignored?).
With a local work size of 2 everything is back to normal. Strange...

The slow uBLAS doesn't need to bring down all the good debugging features globally.

Reported by: Andreas Rost (IHU GmbH)

This test did not return an error code and was hence incorrectly flagged as successful.
What a shame! :-/

Changes:
 - correctly handle an all-zero right hand side
 - iteration count only reflects the true Krylov dimension used
   If orthogonality is lost, the extra basis vectors are not counted.
 - don't run into NaN on residual norm estimator if orthogonality is lost.

Reported-by: Andreas Rost (IHU GmbH)

Savings on my Ivy Bridge laptop are in the range of 15 percent (GCC)
to 50 percent (Clang) in fully optimized mode.
Memory consumption reduced by a similar amount.

Add travis and coveralls support for continuous integration.

Delicate balance with warnings in Clang.

Moved OpenCL-related code for Darwin to ViennaCLCommon.cmake to avoid
code duplication.

These files are not meant to be included by the user -> hide them :-)

Was no longer in use.

Andreas, Denis, and Juraj provided substantial amounts of new code
and should also receive appropriate credits. :-)

Makes it much easier to e.g. only use OpenCL even though CUDA is enabled.
Since this relies on a singleton, the mechanism is not thread-safe.

Migration to Doxygen 1.8.x complete. Resolves #18.

Results in shorter URLs.

A temporary buffer wasn't flushed, still contained old values.

Now deals correctly with very small systems for which the maximum iteration
count is larger than the system size.

Uses temporaries C = trans(B) in order to avoid troubles with corner
cases like A += trans(A).

Enables support for operations like
 C = fabs(A - trans(B))
and is supposed to work with both matrix and vector expressions.

Flags used: -Wall -Wextra -Weverything -pedantic -Werror
-Wno-exit-time-destructors -Wno-global-constructors -Wno-padded -Wno-weak-vtables
-Wno-documentation -Wno-old-style-cast -Wno-switch-enum

For reasons of backwards compatibility and uniformity with other headers.

Flags used: -Wall -Wextra -Weverything -pedantic -Werror
-Wno-exit-time-destructors -Wno-global-constructors -Wno-padded -Wno-weak-vtables