Interop: a host type converter's answer stays with the engine whose converter gave it

[lahma]

TypeResolver's accessor cache is shared between the engines using it, and is partitioned by an
InteropResolutionProfile that reduces a host-installed ITypeConverter to a single bool —
deliberately, because naming the converter itself would pin a host object in a cache that lives for the
process. That bool is what keeps a stock engine and a converter engine apart, and it is why #3522's own
tests pass on this branch. But it puts every engine carrying a converter of its own into one
partition.

Resolution asks that converter a question. IndexerAccessor.TryFindIndexer converts the member name to
the index type of each single-parameter indexer the type offers, and that one answer decides both whether
an IndexerAccessor is built and whether the declared member's PropertyAccessor is handed an
indexerToTry — which ReflectionAccessor.GetValue probes before the declared member. IsShareable
withheld only the first half:

if (accessor is IndexerAccessor && !engine._typeConverterIsDefault)
{
    return false;
}

A PropertyAccessor is not an IndexerAccessor, so the second half passed straight through, in both
directions.

The trace

Measured on 4.x at 953ed25, net472 and net10.0, over a host type carrying both a declared Name
property and a this[string] whose dictionary answers Name and Extra:

order	narrow-converter engine reads bag.Name	wide-converter engine reads bag.Name
narrow first	from-property	from-property ❌ — alone it reads from-indexer
wide first	from-indexer ❌ — alone it reads from-property	from-indexer

Same host type, same member name, and the answer depends on which engine ran first. In the narrow-first
order the stored PropertyAccessor carries no indexerToTry; in the wide-first order it carries one, and
the narrow engine — whose converter refuses that very conversion — is handed an indexer probe it could
never have resolved for itself.

The mirror case is a key only the indexer can answer: the narrow engine stores "no such member" for
bag.Extra, and the wide engine — which has an indexer that answers it — is served that.

The fix

An engine with a converter of its own now neither answers from the shared cache nor adds to it for a type
whose resolution consults that converter. That is a type offering a single-parameter indexer keyed by
anything other than int, which is the one index type TryFindIndexer settles from the member name alone
without asking. The predicate is memoized per resolver and per type, is asked only by an engine whose
converter is not the stock one — so the default configuration pays nothing — and is deliberately blind to
the member filter and to which accessors an indexer carries: answering "yes" too readily costs such an
engine a re-resolution per wrapper, which is exactly what the IndexerAccessor exclusion above already
cost it for the other half of the same decision. Answering "no" wrongly is the bug.

This is the branch-local shape #3560 describes, restricted to the partition 4.x still has. main's
shape — dropping stockTypeConverter from the profile and excluding per type with IsConverterNeutral —
is the v5 refactor, and removing the partition changes an existing default; it is deliberately not here.

Keying the partition on the converter instead was rejected twice over. It would pin the converter and its
engine for the life of the process, which is the reason #3559 gave for keeping an engine's operator
resolutions on the engine rather than in the process-wide table. And weakly keyed — a
ConditionalWeakTable generation counter — it would still grow a cache that never evicts by one entry set
per engine constructed, which is the retention ResolvedAccessorCount exists to bound and
ExtensionMethodEnginesDoNotGrowTheSharedCachePerEngine already pins.

Tests

Four more facts in Jint.Tests/Runtime/InteropAccessorCacheSharingTests, which already owns this bargain:
the two orders of the trace above, the guard that two engines with different converters still share a
type no converter is consulted about (so the withholding stays per type rather than per engine), and the
guard that such engines add nothing to the shared cache per engine — the one the rejected shape would have
broken.

Failing first on unfixed 4.x: 2 of 18 on net10.0, 2 of 18 on net472.

dotnet build -c Release clean; Jint.Tests 6999/0 (net10.0) and 6914/0 (net472),
Jint.Tests.PublicInterface 1820/0 and 1812/0, Jint.Tests.CommonScripts 28/0 both,
Jint.Tests.SourceGenerators 52/0; test262 102,499 passed / 0 failed / 185 skipped of 102,684, matching
the branch control exactly.

No benchmark: the gate is a single bool field test that short-circuits for every engine on the stock
converter, so no default-configuration lane is reached at all. Engines carrying a host converter do lose
cross-engine sharing for indexer-bearing types; that lane was never benchmarked and the cost is bounded by
ObjectWrapper's own per-wrapper descriptor store, but it is the one place worth gating if the interop
suite ever grows a custom-converter row.

Closes #3560