This continues upon #1073 (you may close that one if you prefer in one PR, otherwise I rebase).

I did a local benchmark to see what the cost of the new arity_assign would be in the case we share data. In polars expressions that is every first operation:

Gnuplot not found, using plotters backend
bench_1                 time:   [608.34 us 609.30 us 610.73 us]                    
                        change: [+44.741% +45.228% +45.698%] (p = 0.00 < 0.05)
                        Performance has regressed.
Found 7 outliers among 100 measurements (7.00%)
  2 (2.00%) high mild
  5 (5.00%) high severe

This showed, that in case of shared data, we first do a memcpy and this is ~2x slower than the kernels where we allocate a new memory buffer.

Luckily we can have best of both worlds. :)

This PR branches depending on the shared state:

• in case we share state -> we alloc (and thus choose the fastest path possible)
• in case we own -> we mutate in place (and thus choose the fastest path possible)

I also added a distinction between set_validity and with_validity as we now have mutable data, it is not always logical return a new PrimitiveArray.