Small upgrades to training docs

[mcabbott]

Aims to fix #2325 by mentioning that withgradient accepts a tuple.

And to fix https://github.com/FluxML/Flux.jl/issues/2303 by moving discussion of old-style optimiser structs later.

Also has training doc updates about L1, which should ideally have been in https://github.com/FluxML/Flux.jl/pull/2330 instead. This should be merged afterwards.

[github-actions[bot]]

Once the build has completed, you can preview any updated documentation at this URL: https://fluxml.ai/Flux.jl/previews/PR2331/ in ~20 minutes

[codecov-commenter]

Codecov Report

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

Project coverage is 76.30%. Comparing base (95737ff) to head (9d32b74). Report is 55 commits behind head on master.

:exclamation: Current head 9d32b74 differs from pull request most recent head 6c9cb17. Consider uploading reports for the commit 6c9cb17 to get more accurate results

Additional details and impacted files

@@           Coverage Diff           @@
##           master    #2331   +/-   ##
=======================================
  Coverage   76.30%   76.30%           
=======================================
  Files          31       31           
  Lines        1819     1819           
=======================================
  Hits         1388     1388           
  Misses        431      431           

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