chat: make the wait for vLLM configurable, and give the matrix 20 minutes

[cnygaard]

The supervisor gave up after a hardcoded 900 s with no way to ask for more.

That is generous for a 1.8 GiB checkpoint on an idle card, and too short in two real cases: a large MoE loading from cold storage, and any situation where several servers start at once.

The measurement

Distro matrix on a 16-core box at 5-way concurrency:

never got a server           12     ← of 44 legs
Error in memory profiling     3
still has compute apps        0     ← nothing was actually wrong

The same legs pass solo, and which ones fail moves between runs — ubuntu2604 passed both modes at -n 2 and failed source at -n 5. That is startup outrunning the window, not a distro or mode property.

Change

• glq-chat --ready-timeout SECONDS, defaulting to the existing 900 s so nobody's behaviour changes silently.
• Zero and negative are refused at parse time. "Give up before asking" surfaces as an instant, inexplicable startup failure rather than as the configuration error it is.
• The matrix raises both halves to 20 minutes — its own polling loop and the value it passes to glq-chat, with GLQ_DISTRO_READY_TIMEOUT to override.

That last point is the subtle one: raising either alone just moves which one fires first. The harness waited 10 minutes while the chat gave up at its own 15, so the chat died and the harness's liveness check broke out of the loop early. Both numbers had to move together.

Tests

Three new cases: the default is unchanged, the flag reaches the supervisor, and a non-positive value exits. 81 chat/supervisor tests pass; the generated container script still parses under bash -n, with seq 1 240 and --ready-timeout 1200 both present.