Tor re-tries directory mirrors that it knows are missing microdescriptors

When a microdescriptor for a relay changes, it takes a while to propagate to directory mirrors. In this time, a client can:

1. Download a consensus that references the new microdescriptor
2. Try a directory mirror that has an older consensus, and therefore doesn't have that microdescriptor
3. Repeat 2

This is a particular issue when:

• the client first bootstraps, and the fallback or authority provides a newer consensus than any of its directory mirrors
• the client has been asleep for a while, and its consensus has expired, so it fetches one straight away
• only 1/3 of a client's directory guards has the new consensus

We can fix this by:

• making clients try the same directory mirror for their consensus and microdescriptors
• making clients avoid directory mirrors with missing microdescriptors
• making clients use a fallback when all of their directory mirrors don't have a microdescriptor

This issue affects primary guards and v3 onion services.