test(e2e): run mockable serve chat scenarios on the no-GPU lane (EAI-7484)

[fredespi]

Summary

Move the two mockable serve-chat E2E scenarios off the scarce serial GPU runners onto the GitHub-hosted no-GPU lane, so they gate every PR at no coverage cost. Part of the E2E speedup effort ([EAI-7484]).

The chat scenarios tool definitions accepted (#4) and end-to-end local model (#5) only assert that a served model answers over HTTP — a tools-bearing request is accepted, and a non-empty reply comes back. Neither checks real generation, so neither needs a GPU.

Changes

• Capability-aware serve precondition: a model is served in the background now does a real rocm serve --managed on a host with an AMD GPU (real-serve coverage preserved where hardware exists) and uses the in-process MockServer + a planted managed-service record on a no-GPU host.
• Drop @requires-gpu from chat scenarios 4 and 5 so they also run on the mock lane every PR. Real inference stays covered by the @requires-gpu serve-*-inference scenarios.
• a managed runtime is active becomes a no-op on a no-GPU host (a managed SDK can only install where a GPU family selects wheels; the mock serve needs no runtime). Only affects the two migrated scenarios.
• xfail scoping: the EAI-7423 lemonade-serve xfails for these two ids are scoped to therock_family = "gfx*", so they apply only on a real GPU lemonade host and the scenarios expect-pass on the mock lane instead of XPASS-ing.
• serve-absent-gpu-index-rejected (Replace the openssl CLI with pure-Rust signing & verification #13) stays on GPU (comment updated): on a no-GPU host the GPU-required pre-flight refuses ("no usable AMD GPU") before the --gpu index is validated, so its index-specific rejection can only be observed with a real device present.

Test plan

Linux container gate green:

• cargo clippy --workspace --all-targets -- -D warnings
• workspace + e2e-cucumber lib tests (25 passed)
• cargo xtask e2e mock lane: 4 xfail / 0 XPASS / 0 unexpected; chat Fix CI: prek hooks, headless/flaky test fixes, and drop vendored Codex #4 and Speed up CI #5 now pass on the no-GPU lane.

Real-GPU behaviour (real serve on GPU hosts) is unchanged and will be exercised by the GPU CI lanes.

[rominf]

Reviewed the mock/real lane split, expectation scoping, managed-service setup, and GPU-lane behavior. The mock lane now exercises the intended HTTP contract while preserving real inference coverage, and the focused checks pass.