A qualified destructor-id in an explicit destructor call via `->`/`.`, e.g.
`p->std::vector<int>::~vector()` or `p->A::~A()`, produced an ERROR node.
`qualified_identifier` already accepts a trailing `destructor_name`, but
`qualified_field_identifier` (the member-access path) did not; mirror it there.

The destructor-id in an explicit destructor call may name a class template
specialization, e.g. `p->~vector<int>()` or `p->~shared_ptr<const T>()`.
`destructor_name` only allowed `~ identifier`, so any template-id after `~`
produced an ERROR node. Allow an optional `template_argument_list` and add a
GLR conflict for `destructor_name` to resolve the `<` (template-args vs
less-than) ambiguity. A plain `~Name` is unchanged.

- allow module related statements to appear in preprocessor blocks
- export declarations can now appear anywhere other declarations can

Introduces support for `template for ( init; for-range-decl : expansion-init )`
which is a compile-time way to expand (iterate over):
  - expansion expressions
  - anything destructurable via structured bindings
  - ranges with compile time size

It was sufficient to simply re-use the existing rule for
_for_range_loop_body for the part inside the parentheses, but add the
required sequence of `template` followed by `for`

Related to reflection, annotations use the same syntax as attributes,
but begin with an `=` and support any constant expression. Annotations
and Attributes can't be mixed within [[ ... ]].

This adds support for all the language changes introduced in P2996 which
was recently voted into the upcoming c++26 standard.

This includes:
- reflection expressions: the reflect operator (^^) followed by
  expressions, type descriptors, or the global namespace (::)
- splice specifiers/expressions: [: expression :], which can appear as a
  type, expression, alias descriptor, and more.
- consteval blocks: similar to static_asserts but have been introduced
  specifically to allow constant evaluation side effects to occur at
  specific places in code

Closes #252.

Fixes #301

- _block_item shouldn't contain module import/export/declarations as
  they have to be top-level only
- This allows us to reuse _block_item as the rule for exportable items,
  reducing our state count and parser size quite dramatically

P1102R2 relaxed lambda syntax so that () (empty parameter list) is
optional in more cases.

and move eslint config to package.json

Addresses #224 and properly parses all C++ numbers.

This change allows operators (not just fields/member functions) as dependent
template expressions.

```cpp
auto x = y.template operator()<int>();
```

This syntax is used e.g. to call C++20 lambda expressions with a template head
where the template arguments are not function arguments and therefore cannot be
deduced.

```cpp
[]<typename> () {}.template operator()<int>();
```

This change allows commas inside subscripts.

```cpp
auto x = a[1, 2, 3];
```

From C++23, array subscripts work like arguments to `operator[]` as an n-ary
function.

Before C++23, `operator[]` must be a unary function, but before C++20, it's
possible to use the comma operator inside subscript expressions - and this is
done in libraries that use expression templates to achieve multidimensional
array syntax.

Use of the comma operator in subscript expressions was deprecated in C++20 to
make way for the C++23 change to n-ary `operator[]`.