fix: don't re-walk an already-dereferenced $ref target

[BrianWillows]

Summary

When a $ref is replaced, this.update(newValue) lets traverse descend into the
substituted value:

this.update(newValue)

The same resolved object is substituted at every site pointing at that ref, so
its subtree is walked again on each substitution. When refs nest, every level gets
re-walked once per path that reaches it — a schema where each definition references
the one below it twice is a DAG of n definitions but is traversed as 2ⁿ nodes.

deref() is synchronous, so this blocks the event loop:

definitions	input	before	after
10	1.1 KB	13.5 ms	4.1 ms
18	1.9 KB	579.6 ms	0.8 ms
22	2.3 KB	9,556 ms	1.5 ms
24	2.6 KB	38,250 ms	2.8 ms

A ~2.6 KB schema was enough to occupy a core for 38 seconds. The module's existing
cache only covers refType === 'file' loads, so internal #/definitions/... refs
were re-resolved every time.

Fix

Track which resolved values have already been walked, and skip descending into one
that has been seen before (this.update(newValue, true)).

This is safe because resolved refs are already shared — substituting
#/definitions/d at two sites yields the same object today
(result.properties.a === result.properties.b is true on the current release), so
once it has been dereferenced there is nothing left to do on the next substitution.

Verification

• Existing test suite passes: 41/41.
• Output is byte-for-byte identical to the current implementation across 400
  generated schemas covering nested refs, refs shared between several properties,
  repeated refs, and missing refs (failOnMissing paths).
• The pathological case is now flat: 40 levels resolves in ~1 ms, where 24 levels
  previously took 38 s.

Notes

Circular refs are unaffected — that path (checkLocalCircular / the history
check) is untouched and its tests still pass.

Found and fixed with AI assistance (Claude). Happy to add a regression test that
asserts a deep shared-ref schema resolves in bounded time.

[BrianWillows]

Nudge on this one - checked against master today, still merges cleanly.

The change is confined to one file and adds a visited-set so an already-resolved $ref target isn't re-walked at each substitution site. No signature or option changes.

If you'd prefer the set threaded through explicitly rather than held in the closure, say the word and I'll rework it.