v3.24.1 — slow first tokens from reasoning models no longer time out at 90s

Point Franklin at a big, cache-cold prompt — "synthesize a long document from everything in context" — and a reasoning model can take 60–120s just to emit its first token. Until now Franklin gave up at 90s with a stream timeout, and /retry replayed the same prompt straight into the same wall, so the planning turn looped without ever producing output. The 180s budget meant to cover slow first tokens never applied: the gateway flushes the SSE response headers before the first token, so the wait fell under the shorter stream-idle timer instead of the request timer.

• First-token wait now gets the full request budget. The SSE reader splits two clocks that were tangled into one: time-to-first-token uses the 180s request budget, while the gap between later chunks keeps the tighter 90s idle budget. Slow first tokens are honored; a genuinely stalled mid-stream still aborts fast. Tune each independently with FRANKLIN_MODEL_REQUEST_TIMEOUT_MS and FRANKLIN_MODEL_STREAM_IDLE_TIMEOUT_MS.

Fixes #74. 441/441 local tests pass.