…ng/DOCS.md` (#3286)

Co-authored-by: peterschmidt85 <[email protected]>

* Expand max_duration reference

* Fix typo

SSHTunnel.open() starts the ssh process with -f option, which
causes it to become a daemon (it calls daemon(3) after authentication).

Then, SSHTunnel.close() uses `ssh -O exit` to ask that daemon process
to exit (the processes communicate via the control socket, -S option).

The daemon exits, and then PID 1 should reap it.

Before #3165, PID 1 was bash,
which handles SIGCHLD, properly wait(2)ing children, but in that
pull request exec was added, making `dstack server` PID 1. dstack
does nothing with SIGCHLD (neither handles nor explicitly ignores
it), thus the default disposition (ignore the signal but do not discard
children) leads to an evergrowing number of unreaped zombies.

Fixes: #3291

Minor edits

* Return backend offers from _find_optimal_fleet_with_offers

* Refactor runs.py to package

* Refactor runs into modules

* Implement get_job_plans()

* Return run plan offers wrt fleets

* Implement _get_non_fleet_offers() and _get_job_plan()

* Keep old behavior for dstack offer

* Mark dstack offer as experimental feature

* Return non available offers in run plan

Fixes: #3265

* [Runpod] Require CUDA 12.8+ on the host #3299

* [Runpod] Require CUDA 12.8+ on the host #3299

Updated `gpuhunt` to `0.1.13`

When error logging was added to SSHTunnel.close()[1], it was discovered
that SSHAttach.detach() is called multiple times on detach.
Although this is harmless as long as this method is idempotent (and it is),
this patch ensures that detach() is no-op when called more than one
time, removing ERROR level log noise.

As a side-effect, detaching SSHAttach that reuses SSHTunnel no longer
kills "master" SSHAttach, meaning that one can `dstack attach` multiple
times and detach independently (detaching from the "master" attach still
kicks out all other attaches).

[1]: #3296

* Add SGLang Router Support

* Rename router_config to router

* Rename sglang_workers.jinja2 to router_workers.jinja2

* Resolve SGLang API expose issue

* Resolve model_id based single router to multi router

Test sglang router per service implementation

Test sglang router per service implementation

Test sglang router per service implementation

Test sglang router per service implementation

Test sglang router per service implementation

Test sglang router per service implementation

* Resolve gateway installation command

* Resolved Minor Review Comments

* Resolve comments and kubernetes sglang router intregration

Test gateway package update

Test gateway package update

Test gateway package update

Test gateway package update

Test gateway package update

Resolve rate limits and location issue

Resolve rate limits and location issue

Resolve rate limits and location issue

Resolve all major comments

Resolve all major comments

Resolve kubernetes gateway issue with sglang intregration

* Resolve additional comments

* Pinned sglang-router to 0.2.1

* Fix linting error

---------

Co-authored-by: Bihan  Rana <[email protected]>