Resolves #82 to the extent possible. A bullet-proof fix for 'auto'
requires substantial refactoring and internal changes, which
won't happen prior to ViennaCL 2.0.0.

Introduced by commit 79c21e63
because of a not careful enough fixing of warnings.

Derived from a Radeon HD 6850 with a tuning run.

Obtained on a Radeon HD 5450. Very low-end GPU, profile aims at compatibility
rather than performance.

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

Only requires four kernels per iteration, which is much better than
the Householder version. Implementation follows Algorithm 2.1 in
Walker, Zhou: "A Simpler GMRES" (1994)

Also includes a new example showing the use case.
Resolves #69.

Reported-by: Pushkar Ratnalikar via viennacl-devel

Might have gotten lost during refactoring?

Vector types lead to compilation issues on NVIDIA GPUs with abs(), since
 x = abs(y)
does not compile due to incompatible vector types.

Discusssion here:
https://github.com/viennacl/viennacl-dev/issues/106

No more collisions with GMRES anymore. Resolves #61

Resolves #49

Either vector_base/matrix_base always create shallow copies in their copy-CTOR,
or they create deep copies in their copy-CTOR. The side-effects of shallow
copies can be corrupted data, leading to wrong results. On the contrary,
deep copies are likely to have only poor performance as a side-effect,
which is not as bad. The only use for shallow copies is within
proxy objects, which now have suitable overloads for their copy-CTOR.

Resolves #60

Now avoiding unnecessary temporary buffers.

FFT: Rewrote to fix VS2013 compilation errors

Assuming the complex numbers are in Cartesian form this should be identical in function to the prior function.

Use of 'uint' is rejected by some compilers, e.g. Visual Studio.

Never include this without a separate switch for Apple systems.

Resolves #68. Resolves #76.

v1 -= pow(v1, v2); might be poorly conditioned if v1 is close to 1.0.
This fix lifts the values of v1 to be at least 1.1, hopefully fixing the
repeated issues seen in the nightly tests in the past.

Execution times on CPU are otherwise excessive.
Provide runtime flag for switching sizes later.

About a factor of 20 faster than previous implementation.
I estimate that more microtuning can get another factor of 2.
Higher performance gains will most likely require intrinsics.