Prefix completions with custom python filter

[hendrikmuhs]

this PR combines #283 and #284 and puts custom python filtering on completions on top.

It exposes a set_min_weight method to influence/shortcut DFS traversal during matching. Example for a custom top-5 implementation:

heap=[]
def filter_top_5(completer):
    for m in completer:
        if len(heap) < 5:
            heapq.heappush(heap, m.score)
            yield m
        elif m.score > heap[0]:
            heapq.heappop(heap)
            heapq.heappush(heap, m.score)
            completer.set_min_weight(heap[0])
            yield m

[m.matched_string for m in filter_top_5(d.complete_prefix(key))]

Outdated:

After some discussion the following idea has been dropped.

It allows on the fly filtering and custom weight adjustment from python:

def filter(m):
    accept= True
    weight = 42
    ... # implement some filtering logic
    return accept, weight

d.complete_prefix(key, filter)

[narekgharibyan]

lgtm

[narekgharibyan]

As this method is only available to WeightedTransition, should we prevent calls to it with something like throw std:: logic_error("Not implemented") instead of having a method that is no-op ?

Am good with this as well, just wanna see if this was intentional.

[hendrikmuhs]

I had it like this in an earlier version, but I reverted, because it broke completions on dictionaries without weights (test case should exist, so it is easy to find out).

I will try this once more and make a new PR if it doesn't break this case.

So many changes, at some point I lost track.