Add a top-level `lark-cli skills` meta command group that reads skill
content compiled into the binary at build time via go:embed, so the
content an AI agent reads always matches the running CLI version. It is
an additional, version-consistent source and a fallback for when skills
are not installed or are stale; it does not replace or thin the on-disk
skills.

Commands:
- `skills list` lists every embedded skill with name, description, and
  the frontmatter metadata block.
- `skills list <name>` / `skills list <name>/<sub>` list one directory
  layer (ls-style, no recursion); entries are skill-prefixed paths that
  can be fed straight back to `read`.
- `skills read <name>` prints the skill's SKILL.md and appends a one-line
  tip steering the model to fetch reference files via this command.
- `skills read <name> <path>` / `skills read <name>/<path>` print any
  file under the skill directory (no tip).
- `--json` wraps output in an envelope; on the main SKILL.md it carries a
  separate `guidance` field instead of merging the tip into content.

Implementation:
- skills/ is embedded from the repo-root package main (go:embed cannot
  cross parent dirs) and injected into cmd/skill via a package variable,
  avoiding any change to cmd.Execute's signature.
- internal/skillcontent is pure logic over an injected fs.FS; reference
  reads and directory listings share one path-traversal guard that
  rejects absolute paths, "..", and "..\\" escapes.
- Errors are typed (validation -> exit 2, file_io -> exit 5); rejected
  paths never write anything to stdout.

Risk is set per leaf subcommand (read); the group carries none, matching
the config/service command groups. The commands need no auth (local
embed reads only).

Squashes the review-response iteration on top of the initial skills
command.

- Inject the embedded tree via an init() calling
  cmd.SetEmbeddedSkillContent (no cmd/skill package global, no panic;
  tests inject through the Factory). main.go stays self-contained so the
  single-file preview build (go build ./main.go) still compiles.
- Surface the frontmatter version field in `skills list`.
- Use typed structs for the list and read JSON envelopes.
- Scope the read guidance tip to the skill's own files, route cross-skill
  "../lark-foo/..." references back through `skills read`, and emit it on
  stderr in raw mode (stdout stays byte-identical to SKILL.md) and in a
  guidance field under --json.
- Share one path splitter (skillcontent.SplitArg) between read and list.
- Skip top-level directories without a SKILL.md in List().
- Add usage examples to `skills list` and `skills read`.
- Isolate config state in the command test helper.
- Slim the embed to agent-readable content (SKILL.md + references/, plus
  lark-whiteboard's routes/ and scenes/), excluding machine-resource
  assets/ and scripts/: release binary 26.7 -> 23.4 MB, no change to what
  the commands serve.