fix: compare nodespecs based on ngpu/ncpu/mem first, before price

[mortenpi]

The documentation says that nodespec returns the smallest node, not necessarily the cheapest. This fixes the behavior where nodespec (and therefore submit_job) would pick the wrong node if there is a bigger, but cheaper one available, and clarifies the documentation.

It also makes the output for nodespecs() sorted, but that's an implementation detail, and hence not documented.

[codecov-commenter]

Codecov Report

All modified and coverable lines are covered by tests :white_check_mark:

Project coverage is 40.61%. Comparing base (6799ba1) to head (1964bbf).

Additional details and impacted files

```diff @@ Coverage Diff @@ ## main #49 +/- ## =========================================== - Coverage 78.65% 40.61% -38.05% =========================================== Files 16 16 Lines 2244 2221 -23 =========================================== - Hits 1765 902 -863 - Misses 479 1319 +840 ```

:umbrella: View full report in Codecov by Sentry.
:loudspeaker: Have feedback on the report? Share it here.

[pfitzseb]

1.6 failure seems spurious?

[mortenpi]

1.6 failure seems spurious?

Yep. I missed the doctest failures though.