perf: reduce allocations and redundant work across core, JsonPath, Da…

[SimonCropp]

…taSets and FSharp

Add benchmarks (src/ArgonTests/Benchmarks/) covering every change. All numbers below are probe-measured allocation reductions per operation (net10, Release).

serializer

• indexed loop in GetMatchingConverter: foreach over IList boxed a List enumerator for every value serialized/deserialized
• iterate JsonPropertyCollection/JPropertyKeyedCollection backing lists in hot loops instead of boxing enumerators (SerializeObject, SerializeDynamic, creator path, JObject.WriteTo/GetEnumerator, JContainer copy ctor)
• manual IDictionaryEnumerator use in SerializeDictionary: foreach boxed a DictionaryEntry per entry (-34% dictionary serialize)
• IsInteger(PrimitiveTypeCode) overload: enum-from-integer deserialization boxed the type code and re-derived it reflectively per value (-51%)
• cache string TypeConverters: TypeDescriptor.GetConverter ran per Uri value
• track presence only for required/populated properties (-24% with one [JsonRequired] on the type)
• identity fast path in ToSeparatedCase for already-snake/kebab dictionary keys (-62%)
• HashSet membership + single-pass ForgivingCaseSensitiveFind on the creator path (-42% record deserialize)
• record struct key for the enum cache: Tuple allocated per enum op

reader

• BoxedPrimitives for single-digit reads (-68% single-digit ints)
• span double.TryParse: unquoted floats allocated a dead string per token (-47%)
• skip the intermediate string token for quoted ReadAsDouble/ReadAsBoolean
• trust DateTimeOffset.TryParseExact: dates were parsed twice plus a ToCharArray copy per read (-34%)

writer

• WritePropertyName(string) no longer round-trips through a span back to a fresh string per property
• TryFormat double/float/decimal/Guid/TimeSpan into the pooled writeBuffer instead of intermediate strings (-45..49%)
• JavaScriptUtils writes unescaped runs as spans instead of copying through a rented buffer; SearchValues escape scan for strings >= 16 chars (-46% CPU on clean strings, scalar loop kept for short strings where vector setup loses)

linq / jsonpath

• Descendants() single pointer walk over sibling/parent links instead of nested iterators per depth (22,848 B -> 56 B on a 700-token doc)
• drop the redundant duplicate-name hash when parsing object properties
• typed JValue construction in ReadContentFrom for known token shapes
• static empty converter array in JToken.WriteTo
• bounded cache of parsed JPaths (-50% repeated SelectToken)
• constant-operand fast path in filter evaluation (-29% filter queries)

satellites

• memoize DiscriminatedUnionConverter.CanConvert: FSharpType.IsUnion ran per value for every type once registered (-84% and -94% CPU on plain POCOs)
• cache closed F# list/map delegates: MakeGenericMethod + Invoke ran per call and wrapped exceptions in TargetInvocationException (-83% list deserialize)
• resolve DataTable column names per table instead of per cell (-53%)