New discovery code

[Matt Mackall]

Peter's new discovery code from the sprint is in the tree and I've got
it running on http://selenic.com/hg and /repo.

a) It's really much faster. I've got a test here of a kernel repo that's
over 10k csets behind tip. Old discover takes at least a few dozen round
trips and about a minute, new discovery takes 2 round trips and about a
second. 

b) It's too aggressive for servers that support the known command but
not long args. It defaults to 200 nodes per query, which works out to 8k
bytes, way over the 1k URL limit. This is probably restricted to a
narrow range of people running the post-1.8 default branch, but it bit
me.

-- 
Mathematics is the supreme nostalgia of our time.