feat(reasoning): expose GLM-5.3 thinking effort levels

[DreamFate]

Summary

GLM-5.3 currently exposes only an off/high toggle (zai-toggle / anthropic-manual), based on the conclusion in #1708 that official endpoints only accept a thinking switch. That conclusion is now outdated — Zhipu endpoints have since added effort support.

Verified against open.bigmodel.cn with real keys on 2026-08-17:

Endpoint	Thinking levels	Evidence
/api/anthropic	thinking:{type:'adaptive',effort} — low/medium/high/max all 200 with thinking blocks	adaptive+high already in production use
/api/paas/v4	reasoning_effort accepts low/high/max only	invalid medium returns 400: "请使用 low、high 或 max"
/api/v1 (OpenAI Responses)	reasoning.effort — minimal/low/medium/high/max all natively supported	output length grows monotonically with effort

Changes

Follows the existing GLM-5.2 profile template — no new mechanisms, no UI changes:

• GLM_53_LEVELS: ['off','high'] → ['off','low','medium','high','max']
• Encodings now reuse the three existing generic kinds:
  • anthropic-messages: adaptive-effort (native mapping)
  • openai-completions: zai-thinking-effort (medium degrades to high, same pattern GLM-5.2 uses for level compression)
  • openai-responses: openai-reasoning-effort (native mapping, fixes the profile not matching at all for Responses-mode channels)
• normalizeGlm53Level widened; existing sessions (off/high) migrate losslessly
• Dropped the now-unused zai-toggle / anthropic-manual encoding kinds and their compile branches

Notes

• All three zai channel types are covered automatically: API (zhipu) uses openai-completions; Plan (zhipu-coding) and Team (zhipu-coding-team) share the anthropic-messages path (their default base URL is /api/anthropic, see channel.ts)
• No new channel provider options; /api/v1 users can use the existing generic openai-responses channel type

Verification

• bun run typecheck clean; 489 tests pass (1 pre-existing planning-db failure also on main)
• Runtime check of all three transports: level → normalize → effort mapping matches the table above; glm-5.2 / gpt-5.6 profiles unaffected
• E2E on dev via Plan channel (/api/anthropic): slider "low" produced thinking:{type:'adaptive',effort:'low'} in the request body (captured through a local pass-through proxy); responses normal