Expose CHOLMOD's reciprocal condition number estimate as CHOLMOD.rcond

[ViralBShah]

Fixes #118.

cholmod_rcond was already wrapped in src/solvers/wrappers.jl (both the
_l_ and 32-bit variants), it just had no Julia-level caller, so the
functionality was unreachable. This adds a thin rcond(F::CHOLMOD.Factor) in
the index-type loop next to check_factor, plus a docstring, a public
declaration, an entry in the solvers API page, and tests.

What the estimate is

CHOLMOD computes it from the factor diagonal alone — smallest abs entry over
largest — squared for an LL' factorization, so that cholesky and ldlt of
the same matrix agree:

julia> A = sparse(Diagonal([1.0, 2.0, 4.0]));

julia> CHOLMOD.rcond(cholesky(A)), CHOLMOD.rcond(ldlt(A))
(0.25, 0.25)

(diag of the Cholesky factor is [1, √2, 2], so (1/2)^2; diag of the
LDL' factor is [1, 2, 4], so 1/4 unsquared.)

How good it is

Worth being explicit about in the docstring, since "condition number" invites
the assumption of a norm-based estimate. Over 300 random sparse SPD matrices I
compared it against the true 1 / cond(Matrix(A), 2):

• it never fell below the true value (0 violations), consistent with the
  interlacing of Cholesky pivots between λmin and λmax;
• it was up to 91x optimistic.

So it is exact for diagonal matrices and an upper bound in general — safe for
detecting a badly conditioned or singular factorization, not for measuring
conditioning. The docstring says so.

API choice

I did not overload LinearAlgebra.cond for CHOLMOD.Factor. cond(A, 1)
and cond(A, Inf) already exist for AbstractSparseMatrixCSC via
opnormestinv, and they are norm-based estimates of a different quality; having
cond(F) silently return 1/(a diagonal-ratio estimate) seemed likely to
mislead. Exposing it under its own name matches what the issue asked for
(access to cholmod_rcond) and leaves cond free if a real norm-based estimate
for factorizations is wanted later. Happy to change if you would rather have
cond(F).

Testing

New testset runs for every Tv × Ti combination the CHOLMOD tests already
cover (Float32/Float64 × Int32/Int64), 5 tests each: the exact diagonal
case for both cholesky and ldlt, the 1-by-1 case (1.0), the singular case
(0.0, via check=false), and the bound against 1 / cond(A, 2). The full
test/cholmod.jl passes with no failures, and the new jldoctest passes under
Documenter (confirmed it is actually executed by breaking it once).

🤖 Generated with Claude Code

https://claude.ai/code/session_01L8HenXTVrASo1sBZVGhpDQ

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 92.06%. Comparing base (ba58e74) to head (09be1ea).

Additional details and impacted files

@@           Coverage Diff           @@
##             main     #776   +/-   ##
=======================================
  Coverage   92.06%   92.06%           
=======================================
  Files          12       12           
  Lines        8377     8379    +2     
=======================================
+ Hits         7712     7714    +2     
  Misses        665      665           

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.