rmi by image ID, COMPOSE_PROJECT_NAME interpolation, and Chinese docs

[us]

Closes the follow-ups #77 and #80 left open, and brings the Chinese docs up to date with
what v0.9.0 changed.

rmi by image ID (#79)

mocker images -q | xargs mocker rmi — the standard cleanup — failed, because references
resolved only as names. Resolution now falls back to matching stored image digests by
prefix, but only after every name candidate has been tried, so a real repository name
always wins. The forms accepted are the ones people actually have in hand: a full
sha256:..., the truncated form images -q prints, and a bare hex prefix.

A prefix that matches more than one distinct image is an error naming the count, rather
than deleting whichever came first. rmi is destructive, so guessing is the one thing it
must not do. Every command sharing ImageManager.resolve (inspect, tag, save,
history, push's existence check) gets this for free.

${COMPOSE_PROJECT_NAME} in compose files

The project name was resolved after parsing, so a file interpolating
${COMPOSE_PROJECT_NAME} got an empty string. It is now resolved first — from -p, the
environment, .env, then the directory — and passed into parsing as a variable that wins
over .env and the environment. The file's own name: is one of the sources and cannot be
known before parsing, so the file is parsed again when the resolved name turns out to
differ — what the file reads then equals what its resources are called.

$ mocker compose -p custom-name config | grep -E '^name:|PROJECT'
name: custom-name
      - PROJECT=custom-name

Docs

• COMMANDS.md: rmi documents ID and prefix resolution, including the ambiguity error.
• docs/compose.md: notes that the resolved name is readable as ${COMPOSE_PROJECT_NAME}.
• README.zh-CN.md and docs/zh-CN/compose.md had not caught up with v0.9.0: they still
  advertised network connect/disconnect and described the project name as simply the
  directory name. They now carry the five-step precedence, external:/name: for networks
  and volumes, the project-scoped default network, and the one-network-per-container limit.

Not fixed, now tracked

A compose service can still only join one network (#82). That is the runtime's limit, not
the parser's — attaching a container to two networks fails inside the guest — and it is
also why network connect/disconnect report that they are unsupported.

Verification

swift build clean, swift test 445 tests / 35 suites green (434 before). Against the
real backend: mocker images -q prints sha256:d9e85, and rmi sha256:d9e85, rmi <full digest> and
the bare rmi d9e853e87e55 all remove that image, while rmi sha / rmi sha256: are
rejected rather than matched. compose config interpolates the project name for the directory default,
-p, and the environment.

The matching rule is a pure helper (ImageManager.matchingDigests), so the cases that
matter are unit-tested directly: full digest, bare hex prefix, a prefix shared by two
images (ambiguous, never picked), several tags of one image counting as one, and a prefix
matching nothing.

ref #79, ref #82