If anything, it seems to be *faster* with the default initial
capacities, but the difference is not very significant.

And also expand string concatenations in arguments of
StringBuilder.append, to avoid javac creating unnecessary intermediate
StringBuilders.

StringBuffer is synchronized for thread safety, but Lexer is already not
thread-safe anyway, so this is unnecessary overhead. The other uses of
StringBuffer were local in a single method and so can also be safely
replaced.

This actually uncovered a type mismatch, but I'm not sure whether it's
safer to fix the types or to just remove the dead code (because
apparently nobody in the last 20 years noticed that it's dead...).

Now gets start/end positions directly from nodes where possible instead
of going through getTokenList() and getMapping(). This should also
reduce the amount of SourcePosition objects that need to be created, as
the existing ones from child nodes/tokens are reused.

This saves having to update the firstPopped/lastPopped instance
variables at runtime on every pop call.

They were made protected in 2017 and used for a short time by our
B parser, but this was quickly replaced with a different solution - see:
hhu-stups/probparsers@5fc63a6cb8caaff3d567ebfee1d706230cea00f6

It seems that they are not used anywhere else, so make them private
again so nobody gets any ideas...

This was apparently only ever used by our B parser, and that usage was
removed in 2017:
hhu-stups/probparsers@9a5299c0b865e511e5b4cb22722fe1bab0164cdf