Improve attribute read time in ADIOS2 backend

[franzpoeschel]

Previously, the methods listDatasets, listPaths and listAttributes issued calls to adios2::IO::GetAvailableAttributes and scanned them for the pattern of interest. Two issues:

1. Calling adios2::IO::GetAvailableAttributes is an operation linear in the size of the attributes map since the map is returned by value.
2. Scanning for the pattern of interest was done via a linear search.

The same holds for those methods' usage of adios2::IO::GetAvailableVariables.
This PR solves these issues respectively by:

1. Buffering the aforementioned maps and reloading them only as soon as attributes/variables are modified.
2. Using the fact that std::map<std::string> can be efficiently queried for prefixes, see here.

[ax3l]

Thanks, just some minor comments.

Did you see/measure how this affects read/open performance? Any benchmark to share?

[ax3l]

ping @franzpoeschel just a few slight updates needed for merge :)

[franzpoeschel]

Time measurements, performed with a Series containing 30 iterations as written by the LaserWakefield example of PIConGPU, with 688 variables and 4264 attributes:

Before this PR:

> time ./openSeries 
Opened series stream

real    0m21,834s
user    0m21,666s
sys     0m0,053s

After this PR:

> time ./openSeries 
Opened series stream

real    0m0,516s
user    0m0,462s
sys     0m0,012s

For comparison, when using file-based iteration layout, before this PR:

> time ./openSeries 
Opened series stream%T

real    0m0,713s
user    0m0,694s
sys     0m0,017s

And afterwards:

> time ./openSeries 
Opened series stream%T

real    0m0,167s
user    0m0,157s
sys     0m0,009s

[ax3l]

Really nice, thank you very much!
@williamfgc have you seen this already? :)

[ax3l]

Looks great!
If you like, feel free to update CHANGELOG.md on the way, otherwise I'll do it on release.

[williamfgc]

@ax3l @franzpoeschel cool, great job! Thanks for sharing the good news.