Failure in multi-row INSERTs sets citus.multi_shard_modify_mode to sequential

Citus manually marks `multi_shard_modify_mode` to sequential while executing multi-row INSERTs (https://github.com/citusdata/citus/blob/959629d3f308f5d158ad4f7a4497d45aec61f55c/src/backend/distributed/executor/adaptive_executor.c#L2184-L2188).

And, on failures, it is not cleared until the user re-connects. Steps to reproduce

```SQL
cusred034c9409-vl2:citus onderkalaci$ psql postgres
psql (13beta3)
Type "help" for help.

postgres=# show citus.multi_shard_modify_mode ;
citus.multi_shard_modify_mode
-------------------------------
parallel
(1 row)

postgres=# BEGIN;
BEGIN
insert into users_table(user_id) VALUES (1),(2),(3) RETURNING user_id / 0;
ERROR: division by zero
CONTEXT: while executing command on localhost:9701
postgres=!# ROLLBACK;
ROLLBACK
show citus.multi_shard_modify_mode ;
citus.multi_shard_modify_mode
-------------------------------
sequential
(1 row)

postgres=#
```

Start with the multi-row INSERT handling in src/backend/distributed/executor/adaptive_executor.c at the linked lines, then reproduce the failure with the SQL transaction shown in the issue. Trace how citus.multi_shard_modify_mode is changed and cleaned up after the error; done means it returns to parallel after rollback without reconnecting.

Written by the indexing model from the issue text.