Default reasoning effort to high (was max) — fixes runaway thinking, 500s, empty content

GLM-5.2 template forces Reasoning Effort: Max unless reasoning_effort==high, so hard/long
requests burned the whole token budget thinking (187s, 0 content) -> proxy 500s / timeouts.
Ship chat_template.think.jinja that defaults effort to high (client can still override):
faster, content reliably produced, reasoning + tool calling intact.