swarmexec v1.15.0 — live CPU and memory usage

- Resource usage, measured on the nodes. The manager API only ever knew what
  the scheduler had BOOKED against a node; it could report "8% reserved" about
  a node that was on fire. A new node-local Stats RPC adds the other half,
  and the two stay in separate blocks because they answer different questions.
- Where it shows: a container or service past 70% carries an orange marker in
  the tree, past 90% a red one, naming the resource. A service takes its WORST
  replica, not an average — an average hides the one container about to die.
  The node detail gains a measured block under the reservations. The inspect
  overlay gains a third view: 't' now cycles table, stats and raw JSON.
- Every percentage says what it is a percentage OF: the container's own limit
  when it has one, the node's capacity when it does not. 91% of a 256 MB limit
  and 91% of a 64 GB node are different conversations.
- Memory excludes the page cache, so a container that has merely read a lot of
  files no longer looks like it is about to be OOM-killed. CPU needs two
  readings before it means anything and shows "…" until then, rather than a 0%
  that reads as idle.
- The agent samples only while someone is looking and forgets its readings when
  they stop, so an agent nobody is watching costs nothing.
- Docs: re-rolling the agents without changing the shared secret — including
  the trap that makes it go wrong — is now documented in all five languages.
- Release notes on the site render as written again; a note in prose used to
  come out as one heading per line.

This feature needs the agents redeployed: run `swarmexec init --force`. Until
then the client shows no usage numbers, by design, and says so.