Make replaying reasoning content opt-out

[chengwudi1]

Fixes #6968

Problem

reasoning_content is echoed back on assistant messages so that
OpenAI-compatible reasoning endpoints which require it keep working.
DeepSeek's thinking mode answers HTTP 400 when an assistant turn that
carried reasoning arrives without it (#6296), which is why the replay was
added in the first place.

Endpoints in the opposite position return the property but reject it on
the way back. Groq answers a follow-up request built from such a history
with:

400: 'messages.6' : for 'role:assistant' ...
property 'reasoning_content' is unsupported

The first model invocation succeeds; it is the second one — after
ToolCallingAdvisor has run the tool and rebuilt the request from the
conversation history — that fails. Since the two requirements are
mutually exclusive, no unconditional behavior can satisfy both.

Change

The replay becomes configurable rather than unconditional:

• OpenAiChatOptions.replayReasoningContent defaults to null, which
  keeps today's behavior: the property is sent whenever it is present.
  Setting it to false stops the replay. The assistant turn itself is
  still sent as before — only the reasoning_content property is dropped.
• Exposed as spring.ai.openai.chat.replay-reasoning-content.

Nothing changes for OpenAI itself (the property is only added when the
assistant message carries reasoning content) and nothing changes for
DeepSeek (the default stays on).

Tests

• OpenAiChatModelTests.reasoningContentNotReplayedWhenOptedOut — builds
  a request from an assistant message carrying reasoning with the opt-out
  set, asserts the property is absent, and asserts the assistant content is
  still sent. It fails without the guard in createRequest (verified by
  removing the guard locally: the property reappears).
• OpenAiChatOptions builder coverage: both mutate() and
  combineWith(...) carry the new option.
• OpenAiChatPropertiesTests.chatReplayReasoningContentTest — the property
  binds, reaches toOptions(), and
  chatReplayReasoningContentDefaultsToReplaying pins that it stays unset
  by default.

Verification

./mvnw -pl models/spring-ai-openai -Dmaven.build.cache.enabled=false test
  Tests run: 228, Failures: 0, Errors: 0, Skipped: 0

./mvnw -pl auto-configurations/models/spring-ai-autoconfigure-model-openai \
    -Dmaven.build.cache.enabled=false test
  Tests run: 32, Failures: 0, Errors: 0, Skipped: 0

checkstyle, javaformat and enforcer are clean.

(Both commands are run from the repository root. The build cache is
disabled because it otherwise skips the surefire execution and the build
reports success without running any test.)

Documentation

One row added to the OpenAI chat property table in
spring-ai-docs/.../api/chat/openai-chat.adoc.