Performance improvements for C++

[Rangi42]

In the past few weeks we've been transitioning from C-style manual malloc and free (with over a dozen manually realloced collections, and known memory leaks in RGBASM, including some intentional ones essential to how it works) to C++-style STL generic (containers, algorithms, etc). This has in some cases reduced performance (particularly with older compilers).

Some possibilities for optimization:

• assume instead of assert to allow optimizing out the failed-assertion case.
• reserve an expected capacity for containers before filling them.
• Make sure values are being std::moved instead of copied when possible.
• std::visit uses a table of function pointers in older compilers (try some in this example). We could just be sure to compile the release builds with more recent gcc/clang, but that still wouldn't help the popular Windows builds. We could also write our own solution to visit a std::variant; there are many ways to do so (the linked example has some).
• Bison's C++ parser, using its own variant instead of a union and allowing tokens to have nontrivial constructors, significantly slows things down. We might be able to switch back to a C-style one and add manual allocation of nontrivial token values (plus %destructors).