CSR matrix and docs for PETSc extension

[Iddingsite]

Following #894, I have added support for CSR matrices using SparseMatricesCSR.jl as PETSc uses this format by default. I took some inspirations from GridapPETSc.jl as SparseMatricesCSR.jl was also designed for this package I believe.

Checklist

• Appropriate tests were added
• Any code changes were done in a way that does not break public API
• All documentation related to code changes were updated
• The new code follows the
  contributor guidelines, in particular the SciML Style Guide and
  COLPRAC.
• Any new documentation only uses public API

Additional context

It brings about a 10% speedup in general for large problem (tested from 1000 to 10000 unknowns) compared to CSC format because there is no need to convert from one format to the other.

So things like that work:

using LinearSolve, SparseArrays, SparseMatricesCSR, PETSc, MPI

MPI.Initialized() || MPI.Init()

# ── Build a simple sparse matrix (CSC first, then convert) ───────────────────
n = 100
A_csc = spdiagm(-1 => -ones(n-1), 0 => 2ones(n), 1 => -ones(n-1))  # 1D Laplacian
b = rand(n)

# ── Option 1: SparseMatrixCSR{1} (1-based, default) ──────────────────────────
# Simple but incurs an index-shift on every cold start.
A_csr1 = SparseMatrixCSR(A_csc)

prob = LinearProblem(A_csr1, b)
sol  = solve(prob, PETScAlgorithm(:gmres; pc_type = :ilu))
@show norm(A_csc * sol.u - b)

# ── Option 2: SparseMatrixCSR{0} (0-based) — zero-copy fast path ─────────────
# No index shift needed on cold start; canonical form matches PETSc directly.
csr1    = SparseMatrixCSR(A_csc)
rowptr0 = getrowptr(csr1) .- one(eltype(getrowptr(csr1)))
colval0 = getcolval(csr1) .- one(eltype(getcolval(csr1)))
A_csr0  = SparseMatrixCSR{0}(n, n, rowptr0, colval0, copy(csr1.nzval))

prob = LinearProblem(A_csr0, b)
sol  = solve(prob, PETScAlgorithm(:gmres; pc_type = :ilu))
@show norm(A_csc * sol.u - b)

# ── Option 3: cache + reinit (Case-3 hot path) ───────────────────────────────
# Same sparsity pattern, values change — reuses the KSP.
import SciMLBase

cache = SciMLBase.init(LinearProblem(A_csr0, b), PETScAlgorithm(:gmres))
solve!(cache)

for scale in [1.0, 1.5, 2.0]
    A_new = SparseMatrixCSR{0}(n, n, rowptr0, colval0, copy(csr1.nzval) .* scale)
    SciMLBase.reinit!(cache; A = A_new, b = rand(n))
    solve!(cache)
    @show norm(A_new * cache.u - cache.b)
end

The PR also simplifies some things for the CSC case and fix PETSc version to 0.4.8 because of an important bug fix in PETSc.jl.

Done with Claude's help.

[Iddingsite]

I have added a doc entry because I think it is mostly usable now. What's missing is MPI support. I may give it a try at some point.

[Iddingsite]

I believe Test failures are unrelated to this PR