Use a semantic actions commit-list during parsing / Partial parsing

[ptal]

Currently, the semantic action are called whenever the sub-expression is parsed. It means that if the parsing fails upwards, the result of the semantic action will be dropped. This is terrible if the semantic action has side-effect. It also will avoid to copy arguments passed to rules (when it'll be implemented) and it'll probably optimize the whole parsing step since no semantic rule will be called for nothing.

[ptal]

Probably the easiest technique to implement this is to parse a first time the input without calling semantic actions or constructing the AST and to register the "good path". And to reparse it in a second pass by effectively calling the actions.

[ptal]

This "path" could also be used for recording a partial state where more of the input is needed to continue. It might be less efficient?

[ptal]

A version that does not record path is already implemented by &rule rule, &rule try to parse the input and only construct the AST with rule in case of success.

[ptal]

We need benchmarks for measuring how fast is the recognizing step vs. parsing step. If it is fast enough, I propose to record the path when recognizing. It is doable by compiling the choice as follow:

e1 / e2 --->

let idx = state.replay.pop().unwrap();
match idx {
  0 => [e1],
  1 => [e2]
}

During recognizing we record the corresponding index into a replay index vector. I'm not sure if it's wise to make this the default and to store replay inside the state.

[ptal]

I suggest a simpler version, useful to avoid calling semantic action with side effects on a global environment, by englobing semantic actions in lambda functions.
For instance, e1 e2 > f will produce a function || -> f(arg1(), arg2()) where arg1, arg2 are also functions generated by e1 and e2. Later on, it is possible to pass an environment as parameter, connecting to #10.
It doesn't seem too hard to implement.