Skip creating encoders for unexported fields. They are not participating
in JSON marshaling anyway. This allows using unexported fields of
non-marshalable types in structs.
As a side-effect of this change it's no longer possible to marshal
unexported JSON fields by adding a custom type extenstion. It seems this
is desired behavior since it matches standard library and jsoniter
already disallows `json:"-"` fields from participating in custom
extensions.
Fixes#174.
1. Null values for primitive types no longer clear the original value in
the destination object.
2. Dereference multiple levels of pointers in the destination interface{}
type before unmarshaling into it. This is needed to match stdlib
behavior when working with nested interface{} fields. If the destination
object is a pointer to interface{} then the incoming nil value should
nil out the destination object but keep the reference to that nil value
on its parent object. However if the destination object is an
interface{} value it should set the reference to nil but keep the
original object intact.
3. Correctly handle typed nil decode destinations.
When MarshalJSON was defined on a pointer receiver custom enum type
marshaling/unmarshaling was panicing since the underlying primitive type
was treated as a pointer.
Since method set for pointer receivers includes value receiver methods
we don't really need optionalEncoder and can just use marshalEncoder
directly.
Make sure we do the same thing as stdlib with null booleans by not
touching the original value and discarding the null.
Another somewhat related change is nulling out null interface values in
the original structure. This also matches stdlib behavior.
If the client is using the same jsoniter config with multiple goroutines
it's very likely that few initial operations will encounter a placeholder
encoder/decoder while the real one is being created by another
goroutine. Having a full second sleep seems too conservative, since
encoder/decoder will be created in a very short time. This is very easy
to reproduce in any real environment with a few concurrent requests of
the same type. A few initial requests will have 1s+ response time.
Changing to 10ms should smooth out marshal/unmarshal times for these
initial concurrent requests.
