feat: decoder hardening with proper errors

[Fyzu]

Decoder hardening for untrusted JSON input:

• Cap Struct/Value/ListValue recursion at 100 (matching protobuf's CodedInputStream / JsonFormat limit) so deeply nested input fails with a clean JSONException instead of StackOverflowError.
• Normalize decode errors to fastjson2's native JSONException with position context (offset/line/column): malformed int64/uint64/float/double, timestamp, duration, base64, enum names, and numeric/bool map keys. Keep internal/config errors (missing TypeRegistry, bad target Class, unreachable invariants) as IllegalState/IllegalArgumentException so a config bug never masquerades as bad input. For Any, a missing registry is a config error (IllegalStateException) while a client-submitted unresolvable @type is user-facing (JSONException + position).
• Any @type-first fast path: resolve the content descriptor up front and decode the remaining fields straight off the live reader, skipping the LinkedHashMap buffer + JSON.toJSONString + re-parse the old path required.
• Validate bool map keys ("true"/"false" only) instead of silently coercing any string to false (which also collided keys).
• Guard a null/missing packed-WKT "value" in an Any (previously an uncaught NullPointerException for Timestamp/Duration/BytesValue); google.protobuf.Value with null becomes NullValue, mirroring the top-level field rule.
• Route generated decoders' bytes/enum/numeric+bool map-key parsing through public FieldReader helpers so codegen shares the identical JSONException contract (no inline BASE64.decode / Enum.valueOf / Long.parseLong).

Add BuffJsonHardeningTest (recursion depth, parse-error offsets, Any field-order parity, native error types, packed-WKT null value, bool map keys), validating codegen and runtime paths. Full suite green; decode throughput on par with prior baselines and encoder allocation within committed budgets.