Extending the LLPR module to be compatible with scalar targets where num_subtarget > 1

[SanggyuChong]

As explained in the title. Archetypal case is for DOS learning, and indeed, this is something that I am rolling out for @HowWeiBin's DOS models (and hopefully my future ones 😅 ).

I'm already making this draft PR live so that my work is aware and no time is lost with multiple people working on the same extension.

I will coordinate with @Luthaf and @frostedoyster on some of the TODO's I've already place-marked. A decision has to be made on how to handle the agonistic loss used in DOS training in this case, which I will coordinate with Wei Bin first, then discuss with core mtt devs.

I'm also tagging @ppegolo to take a look so that we can already start thinking about accommodating for more general targets, but this PR will only prioritize this specific case of scalar targets with num_subtargets > 1.

Contributor (creator of pull-request) checklist

• Tests updated (for new features and bugfixes)?
• Documentation updated (for new features)?
• Issue referenced (for PRs that solve an issue)?

Reviewer checklist

• CHANGELOG updated with public API or any other important changes?

[SanggyuChong]

currently waiting on #554 to get merged so that I build on top of it