Just using `check` as a prefix to various keywords would lead to too many
ambiguities in the parser (notably in contracts, where you can have an empty
requires list followed by `check ensures ...` or a `check requires ...`. With
a lookahead of 1, it's difficult for the parser see which rule should apply.
Hence, the lexer will perform itself the lookahead and emit a proper token
(`CHECK_ENSURES` or `CHECK_REQUIRES` respectively).

this is a generalization of `check` vs. `assert` to other annotations. Not all
annotations are relevant though. Currently, there are more annotation nodes in
the AST that can have a flag `{ tp_only_check }` than was deemed useful in
pub/frama-c#25, but said flag can in fact safely be ignored. Moreover, this
commit only add this flag in the AST, but provides no further mean to set it to
true except for the original `check` keyword (i.e. on `AAssert`). The parser
and the behavior of the plugins that can handle the flag will be updated in
subsequent commits

Even if we don't know what to do with the argument of the push, this at least
ensures that push/pop directives stay well parenthesized.

a _noreturn attribute with c11 tag will be rendered as _Noreturn keyword

... so that varinfo and logic_var stay synchronized.

original issue introduced here: b9a1eb43

Avoid warning about missing proto during typechecking.
Also changes a warning into an error, and improves the error message.

In order to avoid an undefined behavior in the C code generated for this macro
(conversion from an absolute integer to a pointer).

or any code annotation that is specifying what the whole next
statement is doing