fix(llmcore): responses payload parity - temperature passthrough, ultra effort (closes #813 A-4/A-5)

[louisss1016]

Fixes the two origin-native llmcore.py payload bugs listed as A-4/A-5 in #813 (root-caused by @GiftedScout — analysis credited, co-author on the commit).

Bug 1 — responses mode drops temperature

_openai_stream's chat_completions branch (and ClaudeSession.raw_ask) both gate on if temperature != 1, but the responses branch never wrote the field: any temperature configured in mykey.py silently no-op'd for every Responses-API endpoint (DeepSeek etc. via api_mode: responses). The official Responses API supports temperature (default 1.0).

Fix: same if temperature != 1 guard appended after the reasoning / max_output_tokens lines. The kimi/moonshot force-to-1 (line 519) and MiniMax clamp (line 520) happen before branch selection, so those endpoints are unaffected.

Bug 2 — reasoning_effort: ultra rejected by the whitelist

BaseSession._enum's valid set ends at 'max'; GPT-5.6's ultra was dropped with a single [WARN] Invalid reasoning_effort 'ultra', ignored. and the payload silently fell back to the endpoint default. Fix: one-line whitelist addition in BaseSession.__init__ (shared by every session type, so this covers chat + responses + native alike).

Note: the Claude-only output_config.effort mapping ({'low','medium','high','xhigh'->'max','max'->'max'}) still doesn't map ultra — it warns there instead. That path is Claude-specific and out of scope; flag in case @cuipengcx90 wants it addressed.

Tests — frontends/tests/test_responses_payload_parity.py (7 tests): monkeypatch llmcore.requests.post and drive a real NativeOAISession.raw_ask with a canned SSE stream — no network, no credentials:

• temperature present (0.5) in responses payload when configured
• temperature omitted at default 1 (payload stays identical for most configs)
• temperature still omitted for kimi model override (forced to 1)
• ultra survives _enum → sess.reasoning_effort == 'ultra' and payload['reasoning']['effort'] == 'ultra'
• high unchanged; bogus still dropped (WARN path intact)
• chat_completions branch shape pinned: flat reasoning_effort key, max_tokens key for non-gpt-5 models

Full suite: 279 passed; the 3 failures (release_qualification, symlink ×2) are the same pre-existing Windows-env baseline.

The fake response implements the context-manager protocol because upstream now calls requests.post under with (TTFT/abort support) — added after @GiftedScout's fork point, so their standalone test script needed adapting.