Reuse factor buffers across re-factorizations (follow-up to #55)

[hbadi]

Follow-up to #55 (merged) — as you suggested, rebased onto current master for a clean history.

What

SparseLU / SparseCholesky / SparseLDL (Double and Complex) now allocate L/U
(and pinv, D) only on the first Factorize call, and Clear() + reuse the buffers
on subsequent calls (same sparsity pattern). This avoids re-allocating the factor storage
on every Refactorize — the point of #55, for Newton / time-stepping loops.

The SparseLU trim (per your note)

You suggested gating the end-of-factorization trim on AutoTrimStorage rather than dropping
it — done. SparseLU.Factorize now wraps the final Resize(0) in
if (CompressedColumnStorage<T>.AutoTrimStorage). So with AutoTrimStorage = false the
L/U buffers stay allocated across Refactorize calls → fully allocation-free reuse for
iterative solvers / time-stepping. I updated the AutoTrimStorage doc to mention this.
(Cholesky/LDL allocate the exact factor size, so no trim is involved there.)

[wo80]

Thanks again, this can be merged as is!

The SparseLU implementation has quite a naive approach to memory management, just doubling the sizes of the storage arrays, in case the number of non-zeros exceed the available space. This tends to leave a lot more memory allocated than actually needed, so I added the resize-to-nonzero-count at the end of the Factorize method. Using AutoTrimStorage to opt out of the default behavior should be the right way to go.

I'll update the nuget package later this week.