I have not tested it, but it is probably slower. The advantage of this is that is follows the HTML 5 parsing spec, so it gives you the same results as a browser when parsing malformed HTML
I have found the lxml API in python leaks a considerable amount of memory, but that was while going through a large amount of data though (Wikipedia dump). This is in contrast to a handful of pages. But that was a while ago, it may be different now.
What do you mean by "leak"? I wrote a scraper that used lxml and a single instance/process ran for month on end in production without any issues. Lxml may have used more memory than some alternatives, but I don't think it was "leaking" memory.
That is likely true. My experience with it was in 2011/2012 ish. Their FAQ[1] claims memory leaks have been fixed over time.