Existing implementation was missing support for a `**`
which was pretty critical. This updates the logic to use an
existing library which already supports `**`.

Co-authored-by: Victor Skvortsov <vds003@gmail.com>
Co-authored-by: Jvst Me <git@jvst.me>

* Add distributed Axolotl and TRL example

* Resolve review comments

* [Docs] Renamed `Fine-tuning` to `Single-node training` for more clarity and consistence

* Remove uv from examples with ngc and remove multi-node example from single node training

* [Examples] Minor improvements regarding TRL and Axolotl

* Update Axolotl Single Node Training Example

---------

Co-authored-by: Bihan  Rana <bihan@Bihans-MacBook-Pro.local>
Co-authored-by: peterschmidt85 <andrey.cheptsov@gmail.com>

* Implement DSTACK_MPI_HOSTFILE

* Document DSTACK_MPI_HOSTFILE

* Implement startup_order

* Implement stop_criteria

* Fix tests

* Test master job waiting for workers

* Exclude new fields for client backward compatibility

* Add reference for startup_order and stop_criteria

* Fix hardcoded enum values in reference

* Minor fixes

Fix apply plan excludes being formed incorrectly,
resulting in broken client-to-server backward
compatibility.

Co-authored-by: Oleg Vavilov <vavilovolegik@gmail.com>

Do not suggest hardcoding regions since it is not
necessary for most backends and can mislead
contributors.