openai_server: use COLI_TEMP as the gateway default temperature

[Blakeolson21]

Summary

• use COLI_TEMP as the gateway default when a request omits temperature
• seed COLI_TEMP from --temp in coli serve, whose gateway runs in-process and never saw the launcher env
• keep an explicit request temperature authoritative
• fall back to the existing 0.7 when COLI_TEMP is absent, malformed, or outside the gateway's 0..2 range

Problem

The launcher publishes --temp through COLI_TEMP, following #509 and #968. The API gateway then replaced that value with its own hard-coded 0.7 whenever request JSON omitted temperature.

The built-in non-GLM chat client omits the field, so coli chat --temp 0.25 silently generated at 0.7. coli serve and coli web had the same mismatch twice over: the gateway ignored an exported COLI_TEMP, and --temp never reached the in-process gateway at all because env_for_engine builds a copy for the engine child only.

Use COLI_TEMP at the point where the gateway chooses a missing request default, and have cmd_serve seed it from --temp for the in-process gateway. Explicit API values still win, and an env value that would not pass the gateway's own temperature validation falls back to 0.7 instead of failing requests that never sent the parameter.

Validation

• python3 -m unittest tests.test_openai_server.TemplateTest.test_coli_temp_is_the_default_for_requests_that_omit_temperature -v
• python3 -m py_compile openai_server.py
• make check

The focused test covers the environment default, an explicit request value of zero, and malformed or out-of-range values (malformed, 5, -1, nan, 1e999), each falling back to 0.7. make check passed all C tests and 446 Python tests, with 57 dependency or platform skips.