fix(proxy): stop cross-account model-default sharing and drop unused warm-up call

Greptile and Claude's PR review both flagged that openai-compatible's derived
family defaults were written but never consumed for model mapping, and that
consuming them would mean guessing sonnet/haiku from arbitrary list position
on an unrelated self-hosted endpoint — not a real signal like it is for Codex.

setDerivedProviderModelDefaults now takes an optional shareAcrossAccounts flag
(default true, so Codex is unaffected) so openai-compatible never populates
the shared derivedByProvider bucket, closing a latent cross-account leak.
The proxy hot-path warm-up call (ensureOpenAICompatibleModelDefaults) added
up to 15s of latency per account's first request for a listing nothing reads,
so it's removed along with the now-dead function; the dashboard's on-demand
/api/models fetch is unaffected.

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>