`--budget`, `--order` and `--seed` are accepted and ignored by `--backend model`, silently

[vyncint]

Good first issue. One small inconsistency, findable and fixable without
knowing the tool.

What 2.0.0 established

tune --backend model writes no run directory, so
#22 made it answer an
--out that the caller took the trouble to type, rather than accept it and do
nothing:

$ launchbound tune reduce-flip --backend model --cc 8.6 --out /tmp/run
note: --out is unused with --backend model — it prints the ranking and writes
      no run directory. `launchbound stage` writes one that `launchbound
      report` can read.

The flags that still go quiet

Two others are accepted and ignored, with nothing said:

$ launchbound tune reduce-flip --backend model --cc 8.6 --order random --seed 42
$ launchbound tune reduce-flip --backend model --cc 8.6 --budget 30m

• --order and --seed are for measured backends — the help says so
  ("Search order for measured backends") — and the model backend ranks the
  whole admitted set analytically, in cost order. There is nothing to order and
  no sampling for a seed to control.
• --budget bounds a measured sweep. The model path measures nothing, so the
  value is parsed and then never consulted.
• --seed is equally inert with --order exhaustive on any backend: an
  exhaustive sweep has no random choice to seed.

Each is silently accepted. Someone who passes --budget 30m --backend model
reasonably believes something is bounded.

Fix

Same shape as #22 — one eprintln! per case, saying what was ignored and what
would honour it. Roughly:

if backend == "model" {
    if budget.is_some() {
        eprintln!("note: --budget is unused with --backend model — nothing is measured, \
                   so there is no sweep to bound");
    }
    if order != "exhaustive" || seed != 0 {
        eprintln!("note: --order and --seed are unused with --backend model — the ranking \
                   is analytic and always in cost order");
    }
} else if order == "exhaustive" && seed != 0 {
    eprintln!("note: --seed has no effect with --order exhaustive");
}

Worth checking whether --allow-unsafe and --reason have the same shape on
any path — those two already reject each other properly on stage, which is
the standard the rest should meet.

Done when

No flag on tune is accepted and silently ignored; each says so once, the way
--out already does.

Getting started

CONTRIBUTING.md has the setup and the DCO sign-off (git commit -s). AI help
is welcome; AI attribution is not. The --out note added in #22 lives in
cmd_tune in crates/launchbound-cli/src/main.rs and is the pattern to copy —
including where it sits relative to the work, so a note about an ignored flag
is printed before the output it does not affect.