async_exec now uses a sans-io strategy

[anarthal]

async_exec now supports partial and total cancellation when the request is pending
Added docs on async_exec's cancellation support
Added detail::exec_fsm and macros for coroutine emulation
Requests are now removed from the multiplexer when an error is encountered

[anarthal]

Today is national's PR bombardment day :)

Quick walkthrough of this PR:

• The main point is trying to split algorithms in async_compose into two parts, to make testing easier. The logic is extracted into a finite-state machine (exec_fsm in this case). The FSM is implemented as a coroutine. When the FSM needs to do something "non-local" (like invoking I/O, communicating with another task, etc), it yields, returning an action (exec_action in this case). The action contains all the information about what should be done. With this setup, exec_op becomes "dumb": it just invokes the FSM and executes the requested action.
• In tests, we invoke the FSM directly, simulating different conditions, like cancellations. Tests are dead simple and reliable, but require a little bit of setup to simulate the right conditions (e.g. they create a multiplexer and invoke the functions that a reader and a writer would).
• Ideally, I'd have used asio::coroutine for the FSM. However, this is implemented internally using __COUNTER__ on Windows, which causes trouble (see ODR violations when using asio::coroutine in header-only code under MSVC mysql#259 for an in-depth explanation). So I've created BOOST_REDIS_YIELD that does almost the same, but without this problem (I already use this everywhere in MySQL).
• exec_action and exec_action_type create together a sort of Rust-like enum, to store what to do next. This is lighter at compile-time and easier to use than std::variant.
• I've added a multiplexer::elem::is_done new status to track when something is done. This simplifies the algorithm.
• The behavior on cancellation should almost match what we already had, with some fixes. For instance, calling reset_cancellation_state is required if we want to support cancellations other than terminal.
• I've used Boost.Core lightweight_test instead of Boost.Test because it's lighter at compile time. I can revert to Boost.Test if you prefer it.
• I've added some extra tests for cancellation, just in case.