TODO: nil, #f, #t are integers
TODO: call/cc broken with or with SCHEME_STACK, also in tinyscheme
TODO: fix .atom names
TODO: |name|? (TOK_STRATOM)
TODO: simpler cell memory management, only one block
TODO: proplists broken due to symbol not a pair anymore
TODO: UNNAMED_OPs have little reason to be in the opcode table?
TODO: (list->vector (string->list (string-copy "abc"))) fails miserably when repeated
TODO: (vector-append #(1 2)) same

- major size optimisations
  - struct layout for both typical CISC and RISC cpus.
  - optional error messages.
  - optional error checking.
  - cell size reduced by not storing both float and int forms of a number,
  - major dependency optimisations (no stdio etc.)
- faster and vastly simpler vector memory management (execution speed
  is roughly twice faster than tinyscheme when optimised for size, and 2.5
  times faster when optimised for speed).
  - no need for consecutive cells
  - more compact and faster vectors.
  - resizable vectors (TODO: not exposed yet).
- faster environment implementation, not hashed by name but by symbol address.
- simpler and faster memory management (no need for sorted free lists).
- much faster call stack implementation, fully supporting continuations,
  while greatly reducing cell thrashing, thus fewer gc's and overall
  much faster execution.
- move caaaar etc. to C, which should effectively save some memory
  (and should also be faster).
- allow full 32 bit size range for strings and vectors, where possible.
- made ports optional.
- made error checking and error strings optional.
- implemented small-integer cache, reducing number of cells allocated.
- ripped out features: dynamic loading, nonstandard C targets.
- fix numerous ISO C violations (mostly _ namespace).
- fix numerous uninitialised value violations.
- fix some bufer overflows (mostly relying on _snprintf terminating with 0).
- fix numerous sign/truncation problems.
- remove limit of 65535 args maximum.
- move ~6kb opcode table into read-only segemnt, saves 6kb rss.
- reduce ~6kb opcode table to 1.2kb.
- symbols use one cell instead of two now.
- subnormals are no longer flushed to zero in some comparisons.
- new procedure: gensym?
- scheme fix: gensym does not intern the symbol.
- scheme fix: properly support symbol names starting with '.'
- scheme improvement: support |identifier| syntax.
- numerous R7RS improvements:
  - many added procedures.
  - aligned existing procedures.
  - multiline string (\whitespace) parsing.
- removed unused 1kb linebuf from scheme context.
- abstract cell pointer types, to quickly experiment with different
  representations.
- use c99 rounding functions.