Make cache policy more useful

[glasser]

Previously, all the logic around updating the cache policy with new
hints was hardcoded inside the cache control plugin. Also,
overallCachePolicy undefined was used to represent two distinct
states: "we don't know anything about cache policy yet" and "definitely
uncacheable". There was a PolicyUpdater class that was part of the CC
plugin that handled the logic of updating a policy.

After this PR, overallCachePolicy is always defined. The former case
is maxAge === undefined and the latter case is maxAge === 0. This
object also has some helpful methods that don't exist on CacheHint:

• restrict and replace applies a CacheHint to it; the former can
  only make the policy more restrictive, whereas the latter overrides
  the current values of whatever fields are defined in its argument
• cacheablePolicy either returns null (if we have no information about
  cacheability or if we know it's not cacheable) or returns a
  CacheHint whose maxAge is positive and whose scope is defined.
  These make PolicyUpdater unnecessary.

Inside a resolver, info.cacheControl.cacheHint also has these methods!
So while you can still do info.cacheControl.setCacheHint(hint) (which
is identical to info.cacheControl.cacheHint.replace(hint)), you also
have the ability to do info.cacheControl.cacheHint.restrict(hint) if
that's what you'd like to do.

Additionally, the maxAge and scope fields on
info.cacheControl.cacheHint now update if you call restrict,
replace, or setCacheHint.

Also, resolvers now have access to info.cacheControl.cacheHintFromType
which it can use to extract cache hints from the schema. (This may be
helpful for a resolver that returns an abstract type, such as
Federation's Query._entities).

Also, extend type @cacheControl(...) directives are now honored
instead of silently ignored. There's also a cache for directive parsing
instead of examining the AST in detail every single time a field is
resolved.

This change is mostly backwards-compatible, unless you accessed
requestContext.overallCachePolicy directly (eg from a plugin). In that
case:

• If you relied on requestContext.overallCachePolicy being undefined
  to mean "not cacheable", you should instead see if
  requestContext.overallCachePolicy.policyIfCacheable() is null.
• If you assigned a CacheHint to requestContext.overallCachePolicy,
  you should pass it to requestContext.overallCachePolicy.replace()
  instead.

[glasser]

I need to double-check that these are acceptable map keys. ie (a) we need to get identical objects repeateedly and (b) the object can't be the same when the schema is updated.

Also maybe this needs to be replaced with an LRU cache since otherwise this is a memory leak as the schema changes over time.

[glasser]

Changing to an LRU cache.