jl_rng_split: fix linear possible correlation between four tasks

In response to the demonstrated ease of constructing linearly correlated
task states, and thus linearly related task RNG outputs, this change
departs from using linear mixing for the compression function, which is
simply dot product in those RNGs. I've shown that we can use any mixing
function that's doubly bijective. This change uses `(2c+1)(2w+1)>>1` to
mix the LCG state and the xoshiro register state and _then_ applies the
PCG output function, which is designed to mask linearity.