Isolate assignment to traces ref in metis code

Hide assignment to this reference behind add_trace and set_traces
function calls.

Improvements to Metis made while investigating bad performance reported
by Anthony, Juliano and Mike.

* Implemented a HOL specific finite model, which knows about
numbers, lists and sets.
* Removed multiple provers: METIS_TAC is now solely based on ordered
resolution.

Latest version of METIS_TAC.

The biggest change is that scheduling provers is done by number of
inferences, not time used, so METIS_TAC is now completely
deterministic. If it works for you on a subgoal, it will work every
time HOL4 is built.

New version of Metis:

+ Uses finite models to guide clause selection in resolution.
+ A Metis test for when HOL is compiled using MLton.

New version of metis with a more powerful prover engine.

Fixed a bug found by Michael Norrish whereby definitional CNF interacted
badly with type reconstruction. This resulted in some invocations of
METIS_TAC returning "error during proof translation", and then failing to
find a proof with types switched on. All fixed now.

Now that STRIP_TAC has changed, METIS_TAC can safely do pre-normalization
splitting using DISJ_CASES_THEN. The only user-visible result of this
should be better performance.

A late change that I forgot to propagate to HOL metis: this makes
METIS_TAC more powerful.

Version 1.3: 17 June 2002

Goals submitted to MESON_TAC when HOL is built .......... 1953
Proved by MESON_TAC ..................................... 1953
Proved by METIS_TAC within 10s .......................... 1946

Between Versions 1.2 and 1.3

* Implemention of definitional CNF to reduce blow-up in number of
clauses (usually caused by nested boolean equivalences).
* Resolution is more robust and efficient, and includes an option
(off by default) for clauses to inherit ordering constraints.

Version 1.2: 19 November 2002

Goals submitted to MESON_TAC when HOL is built .......... 1779
Proved by MESON_TAC ..................................... 1779
Proved by METIS_TAC within 10s .......................... 1774

Between Versions 1.1 and 1.2

* Ground-up reimplementation of the resolution calculus, which now
performs ordered resolution and ordered paramodulation.
* A single entry-point METIS_TAC, which uses heuristics to decide
whether to apply FO_METIS_TAC or HO_METIS_TAC.
* {HO|FO}_METIS_TAC both initially operate in typeless mode, and
automatically try again with types if an error occurs during proof
translation.
* Extensionality axiom now included in HO_METIS_TAC by default.