feat(install): add --yes flag for non-interactive SDK installation (EAI-7956)

[r0x0r]

Summary

Adds a --yes flag to rocm install sdk so an SDK install can proceed non-interactively, and makes overwriting an existing managed ROCm SDK an explicit, opt-in action.

• Outside an interactive terminal, reinstalling over an existing SDK now refuses unless --yes is passed, instead of silently clobbering the active runtime.
• A fresh install (no existing SDK) never prompts and is unaffected.
• --yes also continues to approve required system-package installs (e.g. OpenMPI for vLLM) without asking.

Implementation notes

• therock::install_sdk now returns SdkInstallResult { output, mutated }; callers finalize the runtime (finalize_successful_sdk_install) only when the install actually mutated state, rather than keying off dry_run alone.
• The new yes argument is threaded through the install dispatch, the dry-run renderer, and the runtime-update path.

Testing

• cargo test --workspace / cargo clippy --workspace --all-targets -- -D warnings.
• Two e2e scenarios added in tests/e2e-cucumber/features/runtime_setup.feature covering the user-observable behavior:
  • runtime-install-sdk-overwrite-requires-yes (@requires-gpu) — reinstall without --yes is refused with an error explaining --yes is required. The refusal bails before any download, but the precondition (an active runtime) needs a GPU, so it runs on the GPU lane.
  • runtime-install-sdk-overwrite-with-yes (@requires-gpu @nightly) — with --yes the reinstall overwrites and the runtime stays registered/active. Nightly-gated because it performs a real second SDK install.

Both scenarios are GPU/nightly-gated and therefore exercised on the self-hosted GPU and nightly lanes, not the default PR lane.

[michaelroy-amd]

Current head 792ff7ffce02b4107c3206ac8a10377da0ce4cde is CONFLICTING/DIRTY, so its otherwise-green checks do not validate the merge result. Please rebase onto current main, resolve the install-path conflicts, rerun the required install and Cucumber gates on the rebased head, and then re-request review.

[r0x0r]

Rebased onto current main (now 0a2a239) — the branch is MERGEABLE again.

Conflict resolved (1 file):

• .github/workflows/e2e-selfhosted.yml — main reworked the shared-runtime pre-warm block to the cache-aware cargo xtask e2e-prewarm --channel release path (EAI-8057). That supersedes this branch's edit, which had added --yes to the old inline install sdk guard. I took main's pre-warm block wholesale: the cold install it performs is a fresh install (empty tree), and per the flag's own contract a fresh install never prompts, so no --yes is needed there. The --yes feature itself (in main.rs/therock.rs) merged cleanly.

Also fixed a rebase-induced numbering collision:

• tests/e2e-cucumber/features/runtime_setup.feature — main now uses Scenario numbers 1–5, so this branch's two new scenarios (runtime-install-sdk-overwrite-requires-yes, runtime-install-sdk-overwrite-with-yes) were renumbered from 4/5 to 6/7. The @id: tags were already unique; only the human-facing numbers collided.

Re-ran on the rebased head (local, macOS):

• cargo test -p rocm --bin rocm therock — 77 passed (1 ignored: the ENOSPC disk-fill test)
• cargo test -p rocm --bin rocm sdk_install_approval — the overwrite-only-prompts test passes
• cargo test -p e2e-cucumber --no-run — Cucumber suite compiles green

The GPU/nightly install scenarios (Scenario 6/7, @requires-gpu) run on the self-hosted lane. Leak scan clean; commit signed + DCO. Re-requesting review.