e2e-checkpoint-store: align per-agent env with e2e.yml + harden input

Address trail #706 review findings:

• E2E_CODEX_MODEL: gpt-5.1-codex-mini -> gpt-5.4-mini (match e2e.yml)
• add E2E_GEMINI_MODEL pin (gemini-3.1-flash-lite) so gemini-cli doesn't run an arbitrary default model
• add E2E_CONCURRENT_TEST_LIMIT per-agent (gemini-cli=6, factoryai-droid=1, cursor-cli=2) to avoid rate-limit-induced spurious failures
• route roger-roger through its deterministic task (no --rerun-fails), matching e2e.yml, so a real regression isn't retried and masked
• clarify the agent input description: blank runs all real agents; vogon must be selected explicitly (matches the all-agents list, which excludes vogon as e2e.yml does)
• pass the dispatch input via env in matrix-setup instead of interpolating it into the shell script (type: choice only constrains the UI, not REST dispatch)

The collapsed-matrix copilot-requests/COPILOT_GITHUB_TOKEN scoping is left as-is: it mirrors the established e2e.yml pattern that already runs on every push.