lm32-rtems4.11-gcc doesn't like to compare enums "inherited" from other
enums. We just "manually" code the offset instead.

First, we use the new node_is_op in parse_free. Second, we no longer
have a good error message if op_not "leaks" in test/not.

Since function names are normally not used for variables, we keep them
separate and only allocate an identifier when needed. This will help
with a future migration towards a proper symbol table.

This also reduces the size of the table of predefined identifiers,
slightly accelerating the compiler.

This adds instrumentation to expect() to make test failure non-fatal if
FAIL_ON_ERROR is set to "false", then uses it in the "leak" target.

Also added the corresponding infrastructure to unique.c/unique.h

…ch pool

This is less a regression test bit more an indicator for how adaquate
the structure of the symbol table is.

This changes the compiler to use "struct sym" instead of "const char *"
to refer to variables when communicating with libfpvm. This also
implies that the symbol table moves to "struct sym" and that things
that are not a variable have no identifier associated with them.

The goal is to get rid of "label", which now only serves as a pointer
into the source (or a copy thereof), for diagnostics.

…ruct sym"

We now have a Flickernoise-local "struct sym" that contains the libfpvm
symbol.

This removes the list of variable names from compiler.c and lets
fnp.ids provide the mapping from name to index.