PERF: improve perf of variable window rolling_min/max

[jreback]

The variable window implementation of rolling max/min
is O(n x w) rather than amortized O(n) with a deque
based algorithm

here is a possibility
we could easily use the cython c++ impl of deque. here is a suggestion to use a heap.

[hexgnu]

I took a crack at this using the deque.