lsprof records profile from sub threads, but does not save them.

This can be observed by running the following testcase:

#!/usr/bin/env python
import bzrlib.lsprof
import threading

class thread (threading.Thread):
    def run(self):
        self.a()

    def a (self):
        print 'a'

p = bzrlib.lsprof.BzrProfiler()
p.start()
t = thread()
t.start()
t.join()
p.stop().save('test.callgrind')

I have started working on this, and have a patch that make callgrind work, but I need to get other formats work, and look at tests.