draft: implement the prepare fiber canceling

[palage4a]

When applying of a clusterwide config in a highly loaded cluster, an timeout error often occurs.
In special cases, the error is caused by a timeout at the prepare phase for some instance.

cartridge/cartridge/twophase.lua

Lines 467 to 470 in 97a4629

	local retmap, errmap = pool.map_call(opts.fn_prepare, {upload_id}, {
	uri_list = opts.uri_list,
	timeout = vars.options.validate_config_timeout,
	})

After it happens, we get "locked" instance. (see repro in the linked issue)
It happens because we abort (i.e. release lock) only prepared instances.

cartridge/cartridge/twophase.lua

Lines 521 to 528 in 97a4629

	::abort::
	do
	log.warn('(2PC) %s abort phase...', activity_name)
	local retmap, errmap = pool.map_call(opts.fn_abort, nil,{
	uri_list = abortion_list,
	timeout = vars.options.netbox_call_timeout,
	})

This error handling method works well in cases when the "prepare phase" didn't start or failed before a lock was caught.
But in case of a timeout there is situation when the "prepare phase" has started on an intance, but no answer is followed.
Afterwards, when the procedure has finished, it leads to lock catching and blocks the following configs updates.
For now, to fix this situation, the use of cartridge.twophase.force_reapply method is needed.

But it seems to me, we can resolve this case automatically.
For this, we have to make abort (i.e. release two-phase commit lock) for all instances,
which has started the execution of the "prepare phase" procedure.
For this, we need to do 2 things:

• abort all instances;
• make abort function idempotent.

Implementation of my solution is to stop fiber which performs the preparation procedure.

I didn't forget about

• Tests
• Changelog
• Documentation

Close #2119

[filonenko-mikhail]

Gently closing. No plans to review.