Recovery point manager API

The recovery point manager periodically creates recovery points (see
box.backup.recovery_point.create()) using a pluggable backend. The API
lives on the box.backup.recovery_point table and is available before
box.cfg().

• box.backup.recovery_point.manager_create(name, cfg) -- create a
  named manager and start its background fiber. cfg:
  • cfg.backend (table, required) -- a backend module; the manager
    calls backend.new(cfg.backend_cfg), owns the returned
    instance and tears it down on drop.
  • cfg.backend_cfg (table) -- backend options, passed to new().
  • cfg.create_interval (number) -- seconds between attempts
    (default 600).
  • cfg.timeout (number) -- per-attempt timeout (default 300).
• box.backup.recovery_point.manager_drop(name) -- stop and forget a
  manager (a no-op if it does not exist).
• box.backup.recovery_point.managers -- a table from a manager name
  to its object.
• managers[name]:create_point(opts) -- create a recovery point now.
  opts may carry a label (default <instance>.<manager>.<uuid>)
  and a timeout (default: the manager's per-attempt timeout).
• managers[name]:cfg(cfg) -- reconfigure the manager in place; cfg
  is a partial patch, an omitted option keeps its current value.
• managers[name]:info() -- the manager state: last_point, backend
  (the backend instance's own info) and alerts (the manager's failure
  alert plus the backend's alerts).
• managers[name]:drop() -- same as manager_drop(name).

A backend is a module: new(backend_cfg) returns an instance with
create_point(opts) (returns the point, or nil + error) and optional
drop() and info(). The manager passes opts.label (the point
label, defaulting to <instance>.<manager>.<uuid>) and opts.timeout
(the per-attempt timeout).
Requested by @Serpentian in https://github.com/tarantool/tarantool/commit/29a4f5c82c98a8601f8ff7f7c5db0762ac4ca019.

The issue does not name a documentation file; start by locating the recovery point API entry for box.backup.recovery_point. Document manager_create, manager_drop, managers, the manager methods, and the backend contract using the defaults and lifecycle details in the issue. Done when the public API and backend behavior are documented accurately.