8-bit chars: utf-8 ftw!

64-bit pointers: memory-mapped huge files ftw!

DONE * replace use of null pointer by boolean return value

BAD IDEA! * use memory-mapped file

- don't allocate any std::string for node content since we reuse the original mapped memory

- that also mean we don't need to "unique" strings, in fact it's better to keep things local

=> not working for read-write situations with incremental behavior, which is the "raison d'etre" of IPG!

-> maybe hybrid solution?

DONE * remove isMemoized

* unify Parse and Traverse

* Traverse coroutine design with no temp vector -> true iterator

* memoize choice alternative when

- necessary lookahead is more than 1 ?

- may pass through another node ? => that respect the restriction not to execute actions on backtracked alternatives, but probably inefficient

* hash table should be as compact as possible

- open adressing

- optimize when we rehash

- keep memos for

- long parsing (maybe simply drop shortest one as soon as we hit a non-empty bucket with linear probing?)

- not sure, if hash table are per symbol type, lengths should be similar.

- result that we may reuse shortly because of backtracking

- one hash table for "skip nodes", one for "real nodes"

* --no-lines does not make any sense: line numbers in generated files are only making sense for the developers of IPG... It should be off by default and be enablable (with --debug flag) only in debug builds.

* inline all non-node definitions

* bring back stand-alone functions: much easier to debug and navigate

* optimize end-of-file check using computed maximum length of some nodes (when not infinite)