Performance monitoring test

Performance is an important strength of CuPy.
For better quality assurance, it is nice to have a way to "fixed-point observation" of performance.
By having a baseline set of benchmarks, we can:

* Correctly evaluate performance improvement for each pull-requests
* Detect unexpected performance regression

NumPy has a set of benchmark codes to run on [Airspeed Velocity](https://asv.readthedocs.io/en/latest/).

https://github.com/numpy/numpy/tree/master/benchmarks

I haven't tested them yet, but NumPy benchmark code seems so simple that we can borrow them and then do `s/np/cp/`. It may be a good start point.

https://github.com/numpy/numpy/tree/master/benchmarks/benchmarks

We may however want some additional metrics like:

* memory allocation performance
* comparison with NumPy

Start by reviewing the NumPy benchmark examples linked in benchmarks and benchmarks/benchmarks, along with the Airspeed Velocity documentation. Determine how CuPy benchmarks should establish fixed performance baselines and detect regressions. Done should include a documented benchmark setup that evaluates pull-request performance, with the proposed memory-allocation and NumPy-comparison metrics addressed or scoped.

Written by the indexing model from the issue text.