perf: improve iteration performance

[ritchie46]

Fixed a long standing annoyance. The IntoIterator trait doesn't allow for enough generics so we had to return a Box<dyn Iterator>. Which meant a dynamic function call on every iteration and the iterator is completely opaque to the compiler.

This adds an iter metod directly on ChunkedArray<T> for all T: PolarsDataType. This should be preferred over into_iter at all cost.

Note that it still is a nested iterator (as we iterate over different chunks). So if you can please unpack your nestedness during iteration with:

// when you have null values
for arr ca.downcast_iter() {
   // Non nested iteration so this saves a branch, e.g. fast and completely visible by compiler 
   for value in arr.iter() {
      // do work
   }
}

// when you don't have null values
for arr ca.downcast_iter() {
   // also elides null branch
   for value in arr.values_iter() {
      // do work
   }
}

// when you want to ignore null values
for arr ca.downcast_iter() {
   // fastest way to skip null values.
   for value in arr.non_null_values_iter() {
      // do work
   }
}