Support returning app_iter from renderer.

[lrowe]

Sets response.app_iter = result for renderer results which are not text, bytes or None.

My motivating case here is that rendering very large json responses can lead to 'memory leaks' where python is unable to release the space required for the large string back to the OS. This causes process size to increase as it allocates first 100MB for a large json response, then later for a slightly larger json response, it allocates say 105MB, then 115MB... (These allocations are visible in /proc//smaps.)

My hope is that by using using json serializer that returns an iterable these large string allocations will not be required:

class JSONResult(object):
    def __init__(self):
        self.app_iter = []
        self.write = self.app_iter.append

    @classmethod
    def serializer(cls, value, **kw):
        fp = cls()
        json.dump(value, fp, **kw)
        return fp.app_iter


json_renderer = pyramid.renderers.JSON(serializer=JSONResult.serializer)

[mmerickel]

I think this is great. I'd prefer to explicitly check if the result is iterable prior to assigning .app_iter in order to maintain the previous behavior that else is set to .bytes, however it looks like determining if something is actually iterable is almost intractable. It might be enough to use pyramid.compat.is_nonstr_iter. This is me just bringing it up to think about before merging. Thoughts, @lrowe?

[lrowe]

That could work, I think we'd only need to test hasattr(result, '__iter__') as the str test of pyramid.compat.is_nonstr_iter is already checked first.

As webob.response.Response._body__set actively prevents you from setting non-bytes to response.body, swapping the order will preserve the TypeError which seems a good thing.

I'll try to look into making this change in the next few days.

[mmerickel]

Yay, thank you!