Cast DiscreteGaussian args to RR

[ludopulles]

Currently

params = LWE.Parameters(n=1024, q=257, Xs=ND.Ternary, Xe=ND.DiscreteGaussian(8/sqrt(2*pi)))
LWE.primal_usvp(params)

outputs rop: ≈2^42.4, red: ≈2^42.4, δ: 1.012950, β: 40, d: 665, tag: usvp.
This PR will output rop: ≈2^424.2, red: ≈2^424.2, δ: 1.001567, β: 1413, d: 1621, tag: usvp.

While I refactored NoiseDistribution, I forgot to cast stddev and mean to RR. Oops!

Also: output sparse ternaries as T(p=p, m=m[, n=n]) because we can?

Root cause of the issue: if stddev or mean is not a float type, the PrimalUSVP code cost_gsa checks a predicate lhs <= rhs which is now a Symbolic Expression, causing weird behaviour.

[malb]

Ta!