[Bridges] fix SquareBridge with differing constants in off-diagonal

[odow]

SquareBridge ConstraintFunction getter has a sign error when a symmetrization constraint has a nonzero constant

src/Bridges/Constraint/bridges/SquareBridge.jl, MOI.get(::MOI.ModelLike, ::MOI.ConstraintFunction, ::SquareBridge), line 270.

When entries (i, j) and (j, i) of the square matrix differ by a nonzero
constant c (e.g. f_ij = 3 + x, f_ji = 2 + x), bridge_constraint adds a
constraint (f_ij - f_ji) - c == -c (the constant is moved into the set by
normalize_and_add_constraint). When reconstructing f_ji in the
ConstraintFunction getter, the code subtracts rhs a second time instead of
adding it back, so the recovered f_ji is off by 2c.

Repro (c = 1):

inner = MOI.Utilities.Model{Float64}()
model = MOI.Bridges.Constraint.Square{Float64}(inner)
x = MOI.add_variable(model)
y = MOI.add_variable(model)
f = MOI.Utilities.operate(vcat, Float64, 1.0x .+ 1.0, 1.0x .+ 2.0, 1.0x .+ 3.0, 2.0y)
ci = MOI.add_constraint(model, f, MOI.PositiveSemidefiniteConeSquare(2))
MOI.get(model, MOI.ConstraintFunction(), ci)
# entry (2,1) comes back as `4 + x` instead of the original `2 + x`

This only shows up when the off-diagonal expressions differ by a nonzero
constant (e.g. 1 + x vs 2 + x); the common case of numerically-identical
or purely-linear-difference-with-zero-constant entries doesn't trigger it,
which is presumably why the existing test suite (test/Bridges/Constraint/test_SquareBridge.jl)
didn't catch it.

Fix

Add rhs instead of subtracting it:

f_ji = MOI.Utilities.operate!(+, T, f_ji, rhs)

See issue-001.patch. Verified: the repro above returns the original
function exactly after the fix, and test/Bridges/Constraint/test_SquareBridge.jl
still passes in full (all testsets, 367 @tests total).

Note: issue-003.patch touches a different hunk of the same file
(SquareBridge.jl, the ConstraintPrimal/ConstraintPrimalStart getter and
setter rather than the ConstraintFunction getter); they're independent
fixes and both apply cleanly together.