docs: PRICING TRIPWIRE: Off-peak pricing extended to Chinese public holidays; V4 Pro sunset notice dropped

[github-actions]

The pricing page changed how the peak/off-peak discount window is defined, and the
note that kept DeepSeek V4 Pro alive past its September 14, 2026 sunset date has
been deleted from both the pricing page and the first-call page. Both are things
developers plan cost and model choice around: the first changes which hours get the
half-price rate, the second may change whether deepseek-v4-pro is a safe long-term
default at all. The mirror is now internally inconsistent on the second point --
content/en/updates.md still carries the "we will continue providing V4 Pro after
September 14, 2026" wording that was just stripped from the two pages developers
actually read to pick a model, while the release note for V4.1-Flash
(news260910.md) says V4-Pro is being phased out. Someone should decide whether the
removal means the extension was made permanent or that the phase-out is proceeding.

• Peak hours are now defined as 01:00 - 04:00 and 06:00 - 10:00 UTC, Monday
  through Friday, excluding Chinese public holidays; everything else -- weekends
  and Chinese public holidays in full -- is off-peak. Previously the parenthetical
  only said "all other hours are off-peak", so holidays falling on weekdays were
  ambiguous. Batch/cron users can now reliably schedule around Chinese holiday
  weeks; the zh-cn page states the matching Beijing-time window
  (9:00 - 12:00, 14:00 - 18:00).
• The "we have decided to continue providing API services for DeepSeek V4 Pro after
  September 14, 2026, with billing unchanged" footnote (previously (2) on both
  quick_start/pricing.md and index.md, en and zh-cn) is gone. deepseek-v4-pro
  is still listed in the pricing table at DeepSeek-V4-Pro-0813 with its
  concurrency limit of 500, so the model remains available and priced -- but the
  docs no longer promise it will stay that way. Worth confirming with the docs
  owners before this is treated as "V4 Pro is permanent".
• top_p docs were reworded in three places (api/create-chat-completion.md,
  api/create-response.md, guides/thinking_mode.md, en + zh-cn): "takes effect in
  thinking mode, but values below 0.95 are raised to 0.95" became "only takes effect
  in thinking mode, where the effective range is 0.95-1.0: values below 0.95 are
  treated as 0.95". No behavior change, but "only" and the explicit 0.95-1.0 range
  make it clearer that any top_p you pass in non-thinking mode is discarded.
• Fetch metadata churn: quick_start/pricing.md and two news pages
  (news250325, zh-cn news250120) recovered from the SPA fallback shell and
  fetched real titles again. Going the other way, news251201 (DeepSeek-V3.2
  Release) failed on 2026-09-20 and is now flagged as serving the fallback shell --
  its mirrored content may be stale, and re-checking is cheap.
• Remaining diffs are fetched: date bumps on the FAQ pages and llms-full.txt
  regeneration; no dsh pages changed.