Tune performance of sort algorithm

There are multiple parameters that determine the behavior of the sorting algorithm. Among them:
* which algorithm to use for which `n`s
* how many threads to use, and how to chunk the data
* how many bits to use for the first-step radix, depending on the total number of bits in the data

Some of these options were (crudely) exposed via `dt.options.sort`. We need however the algorithm to be able to pick good defaults, depending on the number of elements to sort, data bit width, and the number of available threads.