rocm serve --engine lemonade: interactive variant selection for owner/repo

Motivation

rocm serve --engine lemonade <owner/repo:variant> serves a Hugging Face GGUF under its canonical name. Today a variant is required — a bare owner/repo, or a partial/ambiguous variant like :Q4 that matches several quantizations, is rejected with an actionable error rather than guessed (to avoid silently serving the wrong quantization). See PR #80.

A nicer UX would be to enumerate the available variants and let the user pick when the variant is missing or ambiguous, instead of erroring.

Proposed approach

• Enumerate a repo's GGUF variants without downloading (Lemonade exposes a variants endpoint used by its own pull flow).
• In the foreground/interactive serve path (which has a TTY, during the serve-plan phase in apps/rocm), present the variants and let the user choose, then proceed with the chosen owner/repo:variant.
• Keep the current non-interactive behavior as the fallback: with no TTY (managed/background serves, scripts, CI), list the available variants in the error message so the user can specify one.

Notes / constraints

• Interactive selection only applies to the foreground path; managed/background serves must keep the clear error.
• Belongs in the apps/rocm foreground layer (TTY + serve plan), not the engine's detached serve_http.
• Invariant to preserve: never silently pick a variant.

Start in the apps/rocm foreground serve-plan layer and trace the existing variant validation and TTY handling. Inspect the variants endpoint used by Lemonade's pull flow, then define the interactive and no-TTY paths. Done means ambiguous or missing variants can be selected interactively in the foreground, while managed and non-interactive serves list choices without silently selecting one.