This commit starts to fill out the support for async in `wit-dylib`. In
doing so this has filled out a number of pieces of functionality
elsewhere and implemented more thorough testing. As a result of this
some issues were discovered which are also fixed in this PR. Notable
changes are:

* Naming of the `task-return` intrinsic is now in a helper in
  `wit-parser` to avoid needing to duplicate that elsewhere.
* A new `roundtrip` test was added to `wit-dylib` which generates
  arbitrary WITs and arbitrary values for the WIT and ensures that the
  values are communicated successfully.
* Composition code shared between the `all` and `roundtrip` tests was
  moved to the `artifacts` crate as a shared dependency of the two.
* In-memory data structures for `wit_func_t` are updated to reflect
  async support.
* Deferred deallocation of bytes is pushed as a responsibility to the
  embedder. Storing values on the stack was getting too tricky and felt
  like a too-clever-by-half implementation. To simplify the generated
  code the responsibility is now on the `wit_dylib_*` implementor to
  deallocate bytes in a deferred fashion.
* The `wit-bindgen` crate is a dependency of `wit-dylib-ffi`, the guest
  bindings of `wit_dylib.h`, for async support.
* The `wasm-tools wit-dylib` subcommand now has configuration, in the
  same manner as `wit-bindgen` for which functions should be async and
  which should be sync.
* The `with_temp_stack` helper was removed as that was also getting too
  tricky to manage and instead functions just always allocate some
  temporary space.
* Existing bindings were updated for async functions when that ABI is
  selected, and new bindings were added for various intrinsics related
  to the lifecycle of async imports/exports.

Perhaps the biggest feature in this commit is the addition of the
`roundtrip` test which is hoped to provide a much more thorough
implementation of testing this crate. Notably it generates an arbitrary
WIT structure *and additionally* generates arbitrary values of this
structure. This has weeded out a number of bugs in the bindings
generator and necessitated various refactorings here. This is intended
to be a crucial part of making this a solid dependency to build on.