scaffold unified LowStorageRKTableau + generic perform_step (2N form)

[singhharsh1708]

Summary

• Introduces LowStorageRKTableau{form,N,T,T2} <: OrdinaryDiffEqConstantCache in a new low_storage_tableaus.jl — a unified tableau struct parameterised on family tag (:two_n, :two_c, …).
• Introduces generic_low_storage_perform_step.jl with plain-dispatch _perform_step_oop! / _perform_step_iip! methods (no @generated, following the pattern from Inline ESDIRK/IMEX perform_step stages; drop @generated and S type-param #3676).
• Converts the 2N family (ORK256, CarpenterKennedy2N54, SHLDDRK64, DGLDDRK73_C, DGLDDRK84_C, DGLDDRK84_F, NDBLSRK124, NDBLSRK134, NDBLSRK144) to use LowStorageRKTableau{:two_n} in place of the deleted LowStorageRK2NConstantCache.
• Thin wrappers in low_storage_rk_perform_step.jl delegate to the generic implementations.

This is the first in a series of stacked PRs (#3683 → #3684 → #3685 → #3686 → #3687 → #3688) that convert all LowStorageRK families family by family.

Test plan

• OOP and IIP paths (williamson on/off) validated against master — step counts and solution values identical.
• CI passes.

[ChrisRackauckas]

family by family would probably be easier, yes.

[ChrisRackauckas]

Steps don't matter for non-adaptive methods, are the solutions identical?

[singhharsh1708]

@ChrisRackauckas Yes, solutions are identical — validated by running each algorithm on a Lorenz problem and comparing the final state vector between the branch and master.

[ChrisRackauckas]

these all already hit the same code, so it's not a substantive change. Do any other caches hit the same form?

[singhharsh1708]

Updated: helper indirection removed — perform_step! bodies are now directly in generic_low_storage_perform_step.jl with no intermediate _perform_step_oop!/_perform_step_iip! layer.

Re 'do any other caches hit the same form': yes — RK46NLConstantCache and SHLDDRK52ConstantCache both use the same 2N update formula (tmp = αᵢ * tmp + dt * f(u,...); u += βᵢ * tmp) but store coefficients as individual scalar fields rather than tuples. They could be folded into LowStorageRKTableau{TwoN} by converting those fields to (A2end, B1, B2end, c2end) tuples — that's planned for PR #6. SHLDDRK_2N has a different step structure (odd/even stage alternation) so it stays separate.

[singhharsh1708]

@ChrisRackauckas i think once its merged we can do that then

[ChrisRackauckas]

see the current comment.

[singhharsh1708]

@ChrisRackauckas ohh i see In previously merged #3590, we unified the 9 low-storage 2N methods under a shared LowStorageRK2NConstantCache, reducing alg_cache methods from 82 → 18
Would it make sense to fold RK46NL into the shared LowStorageRK2NConstantCache path as well? It follows the same 2N formulation; only SHLDDRK52 still seems special because of kshortsize = 2.