Fixes T7130. Fixes T7041. Fixes T7012.

Major change here is partitioning clients. In the Phacility cluster, being able to get a huge pile of instances on a single server -- without needing to run a process per instance -- is desirable.

To accomplish this, just bucket clients by the path they connect with. This will let us set client URIs to /instancename/ and then route connections to a small set of servers. This degrades cleanly in the common case and has no effect on installs which don't do instancing.

Also fix two unrelated issues:

• Fix the timeouts, which were incorrectly initializing in open() (which is called during reconnect, causing them to reset every time). Instead, initialize in the constructor. Cap timeout at 5 minutes.
• Probably fix subscriptions, which were using a property with an object definition. Since this is by-ref, all concrete instances of the object share the same property, so all users would be subscribed to everything. Probably.

• Hit notification status page, saw version bump and instance/path name.
• Saw instance/path name in client and server logs.
• Stopped server, saw reconnects after 2, 4, 16, ... seconds.
• Sent test notification; received test notification.
• Didn't explicitly test the subscription thing but it should be obvious by looking at /notification/status/ shortly after a push.