[exec.when.all] Wording of `impls-for<when_all_t>::get-env` results in inefficient implementations of `get_env()`

[lewissbaker]

The current wording of [exec.when.all] defines the impls-for<when_all_t>::get-env implementation as:

[]<class State, class Rcvr>(auto&&, State& state, const Receiver& rcvr) noexcept {
  return JOIN-ENV(
    MAKE-ENV(get_stop_token, state.stop_src.get_token()), get_env(rcvr));
}

However, this formulation forces the eager evaluation of the get_stop_token query, computing the stop-token which then needs to be curried and stored in the returned environment.

If we imagine many such algorithms using this formulation that are composed (e.g. multiple nested when_all() senders) then we can easily end up with the environment returned from the leaf operation's receiver could include many such stop-tokens, each copied into the final environment.

Instead, what I think we want to do is define an impls-for<Cpo>::query-env(query, state, parent_rcvr) operation rather than get-env. Then we change get_env implementation on basic-receiver to instead return a basic-env that holds a basic-operation* op_; member and then forwards queries to return impls-for<Cpo>::query-env(query, op_->state_, op_->rcvr_);

This would allow environment queries to be computed lazily instead of eagerly computing the values for environment queries.

The definition of impls-for<when_all_t>::query-env would then be:

[]<class Query, class State, class Rcvr>(Query query, State& state, const Receiver& rcvr) noexcept -> decltype(auto) 
    requires std::same_as<Query, get_stop_token_t> ||
             std::tag_invocable<Query, const env_of_t<Rcvr>&>
  if constexpr (std::same_as<Query, get_stop_token_t>) {
    return state.stop_src.get_token();
  } else {
    return tag_invoke(Query, as_const(get_env(rcvr)));
  }
}