Harden Spring AI raw provider diagnostics - #3

Rejecting this PR as-is.

Reason:

• It changes the default Spring AI external raw provider diagnostics from warn to fail.
• That crosses the current library boundary. FastMCP Java should provide the safe provider path, hide managed raw providers it creates, and make external raw-provider risk visible through warning/audit/documentation. It should not fail application startup merely because the application context contains external ToolCallbackProvider beans.
• If an application intentionally bypasses fastMcpSafeToolCallbackProvider and wires raw providers to the model, that is host-application wiring responsibility, not something this SDK should enforce by default.
• fail is still useful as an explicit opt-in for applications that want fail-closed behavior, but the default should remain warn.

Acceptable direction: keep default warn, keep fail and off as explicit modes, and document fail as an opt-in production hardening choice.

Updated in 61210fc to match the requested direction: the default Spring AI external raw provider diagnostics mode remains warn, fail/off stay explicit modes, and docs now describe fail as an opt-in production hardening choice. Verified with mvn test and mvn -Pexamples test using JDK 21.