Fixups for 0.7

capnp-js-base has moved

Fix linking during lite builds

More JSON annotation fixes/improvements

In particular, it should not be necessary to specify a void union member's value explicitly. The discriminator should be sufficient. It should also be permitted to omit default values as long as the descriminator specifies the variant.

(Also make sure that external struct-unions are tested.)

This is particularly needed since JsonValueHandler is re-registered every time a JsonValue field is seen.

Implement annotations to control JSON parsing.

The code was init()ing the union field when the discriminant was seen, but this only works if the field type is a struct or group. Instead we store the field schema and initialize it when we see the value later.

This avoids the need for a weird struct-as-namespace pattern inside json.capnp.

Features:
- Rename any field or enum value for JSON purposes.
- Flatten structs/groups into their parent object, possibly with a prefix.
- Assign a special discriminant field for unions, so that union members can be flattened without ambiguity.

Disallow copy construction/assignment of (Atomic)Refcounted classes

If a multiply-inherited subclass of kj::Refcounted is move-assigned from an object of its other parent class's type, the reference count would get zeroed out. This change makes such mistakes a compile error.

A bunch of issue fixes for 0.7

On Windows, we don't have `fork()` and `exec()`, but rather `spawn()`, which does both in one call. This makes FD inheritance hard -- the parent process has to set up its own FD table to be what it wants the child to receive, then call spawn(), then set things back.

`capnp` does all that, but there was a problem: when it created a pipe, it failed to set the FDs as non-inheritable (Windows equivalent of CLOEXEC). Thus the child would implicitly inherit both ends of the pipe without knowing it. Since it would never close the write end, it would never receive EOF on the read end.

The fix turns out to be simple: create the pipe non-inheritable. We are already separately using `dup2()` to create the copy of the descriptor that we actually want inherited.

Fixes #488.