Abort readers of prepared and then aborted transaction

[alyapunov]

When a transaction is prepared (called commit and yielded, waiting for WAL) it is marked as 'prepared' (untill commit finishes and it becomes 'committed'). Other transactions are allowed to read changes of prepared transaction (depending of RO/RW state and isolation level though). The problem is that if prepared transaction is aborted (because of WAL error) all its readers must be aborted too. This behavior must be identical for vinyl and memtx engines.

Reproduce for insertion

box.cfg{memtx_use_mvcc_engine = true}
local s = box.schema.space.create('test')
s:create_index('pk')
s:create_index('sk', {parts = {{2, 'uint'}}})
local txn_proxy = require('txn_proxy')

local tx1 = txn_proxy.new()
local tx2 = txn_proxy.new()
tx1:begin()
tx2:begin()
tx1('s:replace{1, 1}')
tx2('s:replace{2, 2}')
box.error.injection.set('ERRINJ_WAL_WRITE', true)
local fib = fiber.create(tx1.commit, tx1)
fib:set_joinable(true)
-- select by any index must read {1, 1} and lead to the same behavior.
tx2('s:select{1}')
--tx2('s.index[0]:select{1}')
--tx2('s.index[1]:select{1}')
fib:join()
box.error.injection.set('ERRINJ_WAL_WRITE', false)
tx2:commit() -- must fail, but it does not.

Reproduce for deletion

box.cfg{memtx_use_mvcc_engine = true}
local s = box.schema.space.create('test')
s:create_index('pk')
s:create_index('sk', {parts = {{2, 'uint'}}})
local txn_proxy = require('txn_proxy')

s:replace{1, 1, 1}
local tx1 = txn_proxy.new()
local tx2 = txn_proxy.new()
tx1:begin()
tx2:begin()
tx1('s:delete{1}')
tx2('s:replace{2, 2}')
box.error.injection.set('ERRINJ_WAL_WRITE', true)
fib = fiber.create(tx1.commit, tx1)
fib:set_joinable(true)
-- select by any index must read {} and lead to the same behavior.
tx2('s:select{1}')
--tx2('s.index[0]:select{1}')
--tx2('s.index[1]:select{1}')
fib:join()
box.error.injection.set('ERRINJ_WAL_WRITE', false)
tx2:commit() -- must fail, but it does not.