This reverts commit 5a5a1919.

When picodata process forks, the parent and child are connected by
two unix pipes:
- the fuse (used by the child to detect parent's death);
- the ipc channel (used by the parent to get on-exit messages from
  the child to change its state in the finite-state machine);

The problem was that the parent tried to read the message from the
pipe right after the fork (and before the child's exit). Usually
the pipe was empty at that moment and the parent collected serde's
"failed to fill whole buffer" error instead the real message provided
by the child.

Most of them are already moved to the site repo:
https://git.picodata.io/picodata/picodata/docs/-/tree/97c3e35ec/docs

- `clustering.md` -> https://docs.picodata.io/picodata/clustering/
- `glossary.md` -> https://docs.picodata.io/picodata/glossary/
- `cli.md` -> https://docs.picodata.io/picodata/cli/
- `discovery.md` -> https://docs.picodata.io/picodata/discovery/
- `sharding.md` -> /dev/null
- `topology.md` -> https://docs.picodata.io/picodata/deploy/

Some code in libldap which we don't need requires libresolv,
so we should relax linking rules to let linker drop it.

NB: this commit needs the fixes in tarantool-sys,
so we have to update the git submodule.

Previously we extracted attributes from DDL plan nodes and left
the nodes in arena in the inconsistent state. This approach is
error prone, so now we extract the whole node from the plan arena
before using its attributes.

Previous version of the SQL module hierarchy was overcomplicated.
Current commit simplifies it, so SQL has only two submodules:
router and storage.

Previously we created a very strange error sandwitch when failed
on the storage or router runtimes:
- wrapped traft error into a sbroad one
- wrapped the result again into traft error

Now we wrapp sbroad errors into traft or return traft errors as is.

Now it is possible to drop sharded tables with pico.sql() Lua
function.

Adds support for creation of the sharded memtx spaces via pico.sql().