fix: defer platform detection until first use

[cyphercodes]

Summary

• Defer OkHttp platform detection until Platform.get() is first used.
• Add an isolated JVM regression test proving Class.forName("okhttp3.internal.platform.Jdk9Platform") no longer trips the superclass initialization cycle.

Fixes #9462

Testing

• ./gradlew :okhttp:jvmTest --tests okhttp3.internal.platform.Jdk9PlatformInitializationTest --no-daemon --rerun-tasks
• ./gradlew :okhttp:jvmTest --tests okhttp3.internal.platform.Jdk9PlatformTest --no-daemon --rerun-tasks
• git diff --check

Note: locally I had to point Gradle's daemon JVM criteria at the installed Ubuntu OpenJDK 21 to run the checks; that generated file was restored before committing.

[yschimke]

This feels like a confused line now.

var + lazy + volatile.

[yschimke]

perhaps we make get() responsible for setting it if unset?

[cyphercodes]

Updated the deferred platform initialization follow-up:

• replaced the var + lazy + @Volatile holder with a nullable platform cache
• made get() initialize and cache the platform only when unset
• fixed the missing final newline in the new JDK9 initialization test

Verification:

• ./gradlew --no-daemon --no-configuration-cache -Dorg.gradle.java.installations.auto-download=false -Dorg.gradle.java.installations.paths=/usr/lib/jvm/java-21-openjdk-amd64,/usr/lib/jvm/java-11-openjdk-amd64 :okhttp:spotlessKotlinCheck :okhttp:jvmTest --tests 'okhttp3.internal.platform.Jdk9PlatformInitializationTest'
• git diff --check

[JakeWharton]

Please stop pasting LLM output. If you're not going to bother to write yourself why should we be bothered to read it?