rstr - add blocked rstr build/reuse util

[jeremylt]

Purpose:

Should clean up duplicate allocations

Closes: #N/A

LLM/GenAI Disclosure:

None

[jeremylt]

@hughcars not sure why GitHub won't let me request a review from you, but I think this does what you want in your first commit, but in a way that I can fix the problem in a couple of places.

[jeremylt]

Not sure why only the PETSc examples are showing some sort of double free

[jeremylt]

Ok, all sorted

[zatkins-dev]

there's a few legitimately uncovered code paths in code cov, up to you if we want to make sure we hit them or not.

[zatkins-dev]

@hughcars not sure why GitHub won't let me request a review from you, but I think this does what you want in your first commit, but in a way that I can fix the problem in a couple of places.

It's because they're not a collaborator on the repo

[jeremylt]

Yeah, I'm not worried since this is a reshuffling of existing code instead of new functionality.

[hughcars]

Thanks, this covers what my first commit on #2060 was after, and fixes it in the other backends too. I'll drop that commit and rebase once this is in.

[hughcars]

On Graviton5 (64 ranks, GCC 13), BP1-6 at p = 1-8 against main 6dc68d4, 3 paired rounds: xsmm/blocked +1.9%, sve/blocked +1.9%, opt/blocked +1.2%. Most of it is at p = 1 (+4.0 to +4.7%) and in BP5 (+3.6 to +4.1%). The only significant losses are small, on opt/blocked BP4 (up to 0.5%).