[WIP] Caching for stable expressions with constant arguments v2

Hi!

Here's my work on this so far. The patch is still incomplete, but
works well and passes all regression tests (except plpgsql, which is
not a bug, but will be addressed in the future).

As Tom Lane suggested, I rewrote the whole thing and added a new
CacheExpr type. A "bool *cachable" argument is added to many
expression tree mutation functions to return information whether the
subtree is safe to cache. Hopefully this is more approachable than
"stableconst" :) The caller initializes this to true and callees set
it to false when proven otherwise.

If you build it with -DDEBUG_CACHEEXPR=1 then EXPLAIN VERBOSE will
show CACHE[...] around cached (sub)expressions.
----

In boolean expressions (lists of OR/AND expressions), the cahcable and
un-cachable elements are accumulated into separate lists. A new
boolean expression is then built for the cacheable part and prepended
-- so that it's checked first, in the hopes that the cachable check is
cheaper and short-circuits the rest. Not sure if there are significant
downsides to this.

For example:
Input: stable_expr() AND volatile_expr() AND stable_expr() AND volatile_expr()
Output: CACHE[(stable_expr() AND stable_expr())] AND volatile_expr()
AND volatile_expr()
----

To balance my feature karma, I also took the liberty to clean up some
duplicate code. Previously, all callers of simplify_function() did
their own mutation of argument lists prior to calling. Also, arguments
added later in reorder_function_arguments() and
add_function_defaults() were mutated separately in those functions.
Now all of that is done in one place, in simplify_function().
reorder_function_arguments/add_function_defaults functions don't need
a context argument anymore.
----

Work still to left do:
* Rename eval_const_expressions_mutator to const_expressions_mutator
since it doesn't just evaluate constant expressions anymore?
* Should CacheExpr.subexpr be renamed to "arg" for consistency with others?
* CaseExpr, CoalesceExpr, CoerceViaIO and some others aren't cachable yet
* Clean up some hairy code that I've marked with "XXX"
* Some comments are obsolete/incomplete, should be revised
* Remove the "enable_cacheexpr" GUC, it's useful for testing, but
probably not for real usage
* Currently a cachable expression is forbidden from being a "simple
expression". This probably causes the plpgsql regression. I think the
"real" solution is removing CacheExpr nodes afterwards if the
expression is indeed deemed simple; that gives us the best of both
worlds
---

PS: I still don't know how to get git to produce context diff, so
please excuse my unified diff.

This is also available on my GitHub "cache" branch:
https://github.com/intgr/postgres/commits/cache
But don't rely on that too much, I *will* rebase and alter history.

Regards,
Marti