openai: sanitize NUL/C0 bytes from reasoning_content (carved from #41)

[n-WN]

Summary

• Filter NUL (\x00) and other C0 control bytes from extracted reasoning_content so the multi-turn round-trip stays clean. \t, \n, \r are preserved; DEL is dropped.
• Applies to both extraction sources: the dedicated reasoning_content / reasoning field, and the <think>…</think> fallback inside content.
• Exports sanitize_reasoning_text so other crates (request serializers, session-store writers) can apply the same filter when data arrives via a path other than extract_chat_completions_reasoning.

Motivation

Kimi K2.6 has been observed emitting a stray \x00 inside its own reasoning_content chain-of-thought, then refusing the same string on replay the next turn with:

HTTP 400 "the reasoning_content at position N must be a valid UTF-8 string: string contains \x00"

This is the same fix that shipped in the v16 puffer-kimi binary as part of the honest TB2 run (24/89 PASS) — carved out here to land standalone, since the original #41 branch bundles kimi-oauth credentials that don't belong on master.

Traced to qemu-startup retry-03 failures on the honest Kimi TB2 run, where every retry 400'd on replay until this sanitizer was added to the build.

Test plan

• cargo test --package puffer-provider-openai --test reasoning_extraction — 9 tests pass (7 existing + 2 new sanitizer tests + 1 unit test for sanitize_reasoning_text directly)
• cargo check --package puffer-provider-openai --package puffer-core — clean
• New test strips_nul_byte_from_reasoning_content confirms NUL is removed while \t and \n survive
• New test strips_control_bytes_from_think_block confirms the <think> fallback path is also sanitized
• New test sanitize_preserves_whitespace_and_strips_del documents the exact filter contract