fix: 4d sparse statistics crashed on an undefined accumulator (closes #812)

[sotashimozono]

The first @finch block of _4d_structure_to_dcs initializes d_ijk while the declared accumulator is d_ijkl:

d_ijkl = Scalar(0)
@finch begin
    d_ijk .= 0          # <- the 3d accumulator's name; UndefVarError on first touch

so DCStats could never be built for any rank-4 sparse input. Closes #812.

Reproduction, restated here so the PR stands alone:

using Finch
g = Finch.galley_scheduler()
A = fsprand(20, 20, 20, 20, 0.05)
B = Tensor(Dense(Dense(Element(0.0))), rand(20, 4))
compute(tensordot(lazy(A), lazy(B), ((4,), (1,))); ctx=g)
# before this patch: ERROR: UndefVarError: `d_ijk` not defined in local scope

The same error fires for nested SparseList formats too — this function is the statistics entry for every rank-4 sparse input, whatever the level types.

One-line fix, plus a regression test in the DCStats suite constructing statistics for a rank-4 sparse tensor.

Validated on this branch:

• the reproduction above plans and computes, with values matching a dense reference;
• a Dense(SparseList(SparseList(SparseList(Element)))) rank-4 operand plans, and the output preserves sparsity (stored count = the exact product support).

The generic rank >= 5 statistics path (subset enumeration) is a separate concern and not touched here.