Cleaner disconnect handling.  Better fix for issue #71, and also simplifies the interface and improves robustness.

Eliminate the concept of imbuing messages in favor of the simpler concept of setting a cap table directly on MessageReader / getting one from MessageBuilder.  This eliminates capability-context entirely.  This was made possible by the earlier change which moved capabilities to a separate table rather than storing CapDescriptors inline, but I didn't realize it at the time.

Lift requirement that releaseParams() be called before tailCall() or allowCancellation() -- this is no longer necessary given the protocol simplification that allowed cap descriptors to be interpreted immediately upon receipt.

Extend totalSizeInWords() to also return a count of capabilities, which helps when a separate capability table needs to be allocated as well.  Use this in the RPC system.

Revamp concurrency model, part 1:  EventLoop no longer allows cross-thread event queuing, simplifying many things.  Capability clients are no longer thread-safe, so they don't have to be so const.  In the future, explicit ways to communicate between threads will be re-added, but threads will be treated more like separate vats that just happen to have a particularly fat pipe.  Upcoming:  Remove mutexes.

Implement opt-in cancellation for local calls.  (Previously, cancellation would happen whether or not the callee had opted in.)

Tail calls half-implemented, but I think it's time to delete all the cancellation code before continuing because it will simplify things.

Spin off libraries kj-async and capnp-rpc, so that people who don't want them can avoid their large footprint.

Refactor capability code using fork.  Still too much refcounting, though.  Maybe this calls for a different design for pipelining...

Interface code generation.  Also update method grammar to allow multiple returns, and allow specifying request/response structs instead of parameter lists.