A directory manager can now launch or run in read-only or read-write
mode, depending on whether it manages to acquire a lock on the
filesystem.

In read-write mode, it bootstraps and stores data into the database
as usual.

In read-only mode, it assumes that another process will be updating
the database, and so it only loads it periodically, looking for new
information.  However, it also tries to acquire the lock itself, so
it can enter read-write mode.

This function loads a current bootstrapped directory from disk, if
it can.

In C tor, this saved a whole lot of RAM.

This way we won't even try sending the CREATE2 cell unless the
Ed25519 identity we know for the channel matches the provided
CircTarget.

This lets us reinstate the code in dirclient that retired circuits
depending on the error type.

Also, remove references to DataStream, and comment on the future of
the module.

This is consistent with the other pieces of tor-proto, which do not
handle timeouts on their own.  It also lets us remove tor-rtcompat
as a dependency from tor-proto, and simplify some of the test cases
to use async_test.

This commit unindents a lot of test code; use git's "-b" flag to
read the parts that matter.

Now we don't need runtime-specific stuff in tor-chanmgr.

Now other crates don't need any 'ifdef tokio' code, since there
are wrappers that implement 'futures' right.

Technically, the 'futures' traits are in some ways less good than
the tokio ones, but we need a consistent API if we want to support
WASM someday and keep support for async_std.  I'd rather hold out
hope for a future version of futures::io working like tokio than to
fix ourselves into the tokioverse forever.

At this point I'm going to pause writing tor-proto tests for a
little while and think about refactoring that could make these tests
simpler.  I'm doing way too much copy-paste here.

dalek-crypto is stuck on rand_core 0.5.1, so we've been stuck too.
This commit introduces a compatibility module so that we can wrap
new rand_core instances to make them backward compatible.

Closes #106

This includes revising the pick_weighted() function to not panic
if its inputs are above 2^64 in total.