Add an example to show how to debug performance with loggers.

[tcojean]

This PR adds an example mostly taken from the benchmarks to show how to use loggers to debug performance and residual convergence in Ginkgo.

The approach taken is to pass in as argument an optional file (and executor) which is the input matrix file in matrix market format.
In the code itself, at the beginning of the main functions, all important information for configuring the benchmark can be changed manually. The only other place the user may need to change is the preconditioner's settings when using a more complex one, in the solver factory declaration.

The benchmark has three segments:

1. warmup
2. timing
3. extract residual/operation timing information which can have significant overhead.

Tell me if some other information should be shown.

Example of output (reference, simple example matrix):

Generate time (ns): 2303
Apply time (ns): 189355
Residual_norm: 2.10788e-15
Generate operations times (ns):
Apply operations times (ns):
        allocate: 31264
        cg::initialize#8: 1711
        cg::step_1#5: 23430
        cg::step_2#7: 25184
        copy: 12082
        csr::advanced_spmv#5: 40449
        csr::spmv#3: 35953
        dense::compute_dot#3: 21023
        dense::compute_norm2#2: 35839
        free: 20292
        residual_norm_reduction::residual_norm_reduction#9: 8183
Recurrent Residual Norms:
[
        4.3589
        2.30455
        1.46771
        0.984875
        0.741833
        0.513623
        0.384165
        0.316439
        0.227709
        0.170312
        0.0973722
        0.0616831
        0.0454123
        0.031953
        0.0161606
        0.00657015
        0.00264367
        0.000858809
        0.000286461
        1.64195e-15
];
True Residual Norms:
[
        4.3589
        2.30455
        1.46771
        0.984875
        0.741833
        0.513623
        0.384165
        0.316439
        0.227709
        0.170312
        0.0973722
        0.0616831
        0.0454123
        0.031953
        0.0161606
        0.00657015
        0.00264367
        0.000858809
        0.000286461
        2.10788e-15
];

[codecov]

Codecov Report

Merging #360 into develop will increase coverage by 0.79%.
The diff coverage is n/a.

@@             Coverage Diff             @@
##           develop     #360      +/-   ##
===========================================
+ Coverage    89.23%   90.02%   +0.79%     
===========================================
  Files          286      136     -150     
  Lines        21882     6075   -15807     
===========================================
- Hits         19526     5469   -14057     
+ Misses        2356      606    -1750

Impacted Files	Coverage Δ
core/base/iterator_factory.hpp	0% <0%> (-100%)	⬇️
core/solver/gmres.cpp	0% <0%> (-100%)	⬇️
omp/components/format_conversion.hpp	0% <0%> (-100%)	⬇️
include/ginkgo/core/base/version.hpp	0% <0%> (-100%)	⬇️
include/ginkgo/core/base/exception.hpp	0% <0%> (-100%)	⬇️
reference/components/format_conversion.hpp	0% <0%> (-100%)	⬇️
core/base/extended_float.hpp	0% <0%> (-80.22%)	⬇️
include/ginkgo/core/log/logger.hpp	20% <0%> (-77.62%)	⬇️
include/ginkgo/core/base/utils.hpp	30% <0%> (-70%)	⬇️
include/ginkgo/core/base/array.hpp	41.3% <0%> (-55.25%)	⬇️
... and 190 more

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 243234d...13da137. Read the comment docs.

[pratikvn]

LGTM!

[thoasm]

It looks really good.
But I would prefer if you would mention the namespace gko:: for every lend and clone (and similar functions, but I think here, it was mostly these two).

[thoasm]

LGTM!