drivers: stop the ttl fiber in rw mode as well

[maksimuimin]

Closes #262

Problem

fifottl/limfifottl stop the ttl fiber through an unbuffered fiber.channel() that the fiber reads only in the box.info.ro branch. So while the instance is rw:

• tube.raw:stop() blocks the caller forever;
• tube:drop() never stops the fiber (fifottl has no drop()), leaking one fiber per dropped tube;
• utubettl (buffered channel) returns from stop(), but the fiber keeps running and crashes on the dropped space.

Repro and details in #262.

Fix

The fiber now works while it is the fiber registered in self.fiber:

• stop() deregisters the fiber, wakes it up and joins it, so it is guaranteed to be gone when stop()/drop() return, in rw and ro mode alike;
• the iteration re-checks the registration before its long cond:wait(), so a stop request that arrives mid-iteration is not lost;
• start() uses fiber.new() so the fiber cannot observe itself as unregistered before self.fiber is assigned;
• fifottl got a drop() that stops the fiber before dropping the space;
• limfifottl forwards start()/stop()/drop() to the parent object, otherwise self.fiber = nil in the wrapper would shadow the parent's registration and the fiber would never stop;
• sync_chan is removed, the ro branch waits on the same cond with a 100 ms timeout as before.

New test t/260-ttl-fiber-stop.t covers stop/start/drop for all three drivers. Full suite passes locally (Tarantool 3.8, memtx and vinyl).