Further optimizations for province automapping

[IhateTrains]

Notable changes:

• Drastically reduced the number of isProvinceMapped calls (which turned out to be very expensive).
• Reduced the number of locks when registering source-target province matches, thus increasing the gains from parallelism.
• Included https://github.com/martinus/robin-hood-hashing for its faster (about 2x in my test runs) implementation of unordered_map. EDIT: replaced with https://github.com/greg7mdp/gtl
• Made Province struct a class because its instances are quite big, and added some encapsulation.
• Moved some automapping-related code from ImageFrame to Automapper, which seems a better place for it.
• Some micro-optimizations like making classes final.
• A source map pixel is no longer skipped if (sourcePixel.x + sourcePixel.y) % 2 == 0. The code is about 32x faster, despite checking all source map pixels instead of only half of them, so it's no longer needed to reduce the mapping accuracy.

Logged time measurements on Ryzen 7 7700, with I:R Terra Indomita to CK3 Rajas of Asia mappings being generated.

before (measured on an older version of Rajas of Asia, but shouldn't matter much):

Elapsed time (sec): 34.02
Elapsed time (sec): 76.0615
Elapsed time (sec): 70.8512
Elapsed time (sec): 71.7781
Elapsed time (sec): 69.4187

after (see new updated in comment below):

Elapsed time (sec): 9.58116
Elapsed time (sec): 2.19897
Elapsed time (sec): 2.26537
Elapsed time (sec): 2.25273
Elapsed time (sec): 2.23148

The first run always takes a longer time, because a point-province map is being initialized. It's reused in subsequent runs.

[IhateTrains]

Made some changes because I realized that already mapped provinces can't be excluded from tgtPointToWaterProvinceMap and tgtPointToLandProvinceMap if the maps are to be used in subsequent runs, in which the previously mapped provinces can be unmapped.

Also replaced https://github.com/martinus/robin-hood-hashing with https://github.com/greg7mdp/gtl. The previous library was archived, and the new one also provides a custom hashset implementation.

New time measurements:

Elapsed time (sec): 9.60111
Elapsed time (sec): 2.93702
Elapsed time (sec): 2.92947
Elapsed time (sec): 2.90784
Elapsed time (sec): 2.99318

The PR's ready for review again.

[Zemurin]

It certainly appears ok, didn't find anything dubious. Hope everything else works as well. :)