Fix binding- and assignment-expression related typing issues.

Simpler solutions are better, so the typechecker is back to normal.
Instead we simply make sure that assignment expression is never
nested under another expression, and binding expressions may only
be nested under LogicalAnd-expressions. If only I knew why I thought
type shenanigans were a good idea in the first place...

Initial casting implementation

Explicit casts can be performed with the syntax 'cast<type>(input)'
and implicit casts can be performed with the syntax 'cast(input)'
where the output type is determined by inference.

To prevent casting shenanigans we only allow casting of primitive
types and of types to themselves (essentially creating a copy).

Put expression types of procedures in type table

Before we annotated the AST. But that is silly in hindsight:
certain operations have a completely different meaning when applied
in a polymorphic context (i.e. the field index determined by a
select expression).

Rewrote typing to use indices.

Currently it is slower than before, because we do a HashMap lookup
followed up by actually using the index. But it serves as the basis
for a faster type inferencer.

The main goal, however, is to fix the manner in which polymorph
types are determined. The typing queue of functions still needs to
correctly write this data to the type table.

Partial reimplementation of compiler and TypeTable

Every type symbol and its possible polymorphic variables are now
parsed up front and put in the SymbolTable, doing away with the
complicated and error-prone NamespacedIdentifier. Still pending
changes to parts of the compiler and the runtime, so not working
at the moment.

Initial union implementation

Still untested, but: make union declaration implicit (so no longer
inferred from 'enum' use), and initial implementation of type
inference.

Preparatory work for union literals

Contains horrible parsing hacks that transmute function calls and
enum literals to union literals if appropriate. Pending the
implementation of the tokenizer the AST can be constructed more
neatly.

Reimplement namespaced identifier to support polymorphic args

This is somewhat of a temporary hack, as a namespaced identifier should
not really refer to types or polymorphic arguments. But we need a
tokenizer and a prepass to properly distinguish identifiers from types.
So it works, but error messages may be cryptic.