Publish the ModelExpress guide and discovery links

Description

Add the public ModelExpress guide for self-managed compute planes and register it
in both discovery locations.

This covers the parts that do not depend on measured results. Scale-out results,
verification evidence, and fabric validation arrive in the follow-up
documentation task.

Definition of Done

• The guide explains when peer-to-peer weight distribution helps and when it does
  not, so readers can decide whether to enable it.
• Supported versions, worker image contents, NIXL and RDMA requirements, RBAC,
  secrets, and provider-specific network settings are explicit.
• Prerequisites, configuration, rollout, and cleanup are covered.
• The guide states that ModelExpress is optional, opt-in, and installed outside
  the compute-plane stack.
• The page is registered in fern/versions/dev.yml and linked from
  docs/user/optional-enhancements.md in the same change.
• Content changes are confined to docs/user, plus the navigation entry in
  fern/versions/dev.yml noted above. The frozen docs/v0.5, docs/v0.6.0, and
  docs/v0.6.1 trees are unchanged.
• tools/ci/check-docs and fern check pass.

Resources

Related documentation request: #108
Comparable page: docs/user/cluster-management/container-cache.md

The guide deliberately carries no benchmark numbers. Measured figures depend on
model size, hub, fabric, and worker count, and a number quoted without those
conditions is misleading. The guide instead states the conditions that decide the
outcome and tells readers how to measure their own.

Part of #1113.

Start with docs/user/cluster-management/container-cache.md and docs/user/optional-enhancements.md, then inspect fern/versions/dev.yml for navigation conventions. Add the ModelExpress guide under docs/user, link and register it, and run tools/ci/check-docs and fern check; done means the listed prerequisites, configuration, rollout, cleanup, and opt-in boundaries are documented without benchmark claims.