Propagate ParameterArrayAssignments to SCC subsystems

[KristofferC]

This PR is on top of #4697 because it is kind of the same idea. The difference here is that we need to do a little bit of bookkeeping to keep track of the extra cachesyms that each subsystem appends.

This takes ODEProblem construction for a somewhat large model from

6.921406 seconds (108.02 M allocations: 4.142 GiB, 13.60% gc time)

to

  5.740030 seconds (93.43 M allocations: 3.779 GiB, 10.40% gc time)

AI-generated commit message:

---

🤖

The parameter array-variable decomposition computed by
compute_array_variable_buffer_idxs is, like the parameter reordering, a pure
function of the shared parameter layout. Compute the param-only decomposition
once and propagate it to every SCC subsystem.

The per-SCC callers (generate_rhs, generate_update_A, generate_update_b)
append cachesyms as a suffix of the parameter slice, so the cache stores only
the parameter buffers and build_function_wrapper decomposes the (small)
cachesym tail on top of the cached result, shifting its slice-relative buffer
indices past the parameter buffers. The callers pass n_param_buffers (the
number of reorder_parameters buffers before the cachesyms) to mark the
boundary. cachesyms are prior-SCC unknowns / CSE temporaries -- never array
parameters -- so their array-variable keyset is disjoint from the parameters',
making the split identical to decomposing the whole slice.

---