Parsing only the top 3 HSP of each Hit, without parsing the whole file, would require you to write your own custom XML parser. Biopython's NCBIXML does not do this.
However, if it's speed improvement you're looking for, you could try the new SearchIO submodule (http://biopython.org/wiki/SearchIO). It has a new BLAST XML parser that's supposed to be faster than the old NCBIXML parser. The old parser relies on a pure-Python XML parser, while the new one in SearchIO uses cElementTree
whenever possible.
The submodule is still new and experimental, so there might still be some changes before it hits an official release. If you're interested, there's also a draft tutorial here: http://bow.web.id/biopython/Tutorial.html#htoc96.