fix: JNI local reference cleanup in JVMClasses::with_env

[0lai0]

Which issue does this PR close?

Closes #4172

Rationale for this change

JVMClasses::with_env attaches native worker threads to the JVM and then runs JNI call-site closures directly. Executor worker threads can stay attached for the JVM lifetime, so JNI local references created inside those closures may accumulate instead of being released at each call boundary.

Wrapping each with_env closure in a JNI local frame gives all existing call sites a consistent push/pop boundary and prevents long-running executor threads from retaining local references across batches.

What changes are included in this PR?

• Wrap JVMClasses::with_env closures with Env::with_local_frame.
  • Preserve the existing E: From<CometError> caller contract by adapting local-frame JNI errors internally.
  • Document that callers must not return JNI local references from with_env.
  • Audited existing callers: scan, shuffle scan, scalar subquery, memory pools, and parquet encryption convert JNI values to Rust-owned values before returning.

How are these changes tested?

• cargo fmt -p datafusion-comet-jni-bridge
• cargo check -p datafusion-comet-jni-bridge
• cargo test -p datafusion-comet-jni-bridge
• cargo check -p datafusion-comet --all-targets
• cargo clippy -p datafusion-comet-jni-bridge --all-targets -- -D warnings
• cargo clippy -p datafusion-comet --lib -- -D warnings
• JAVA_TOOL_OPTIONS="-Xcheck:jni" ./mvnw test -Dtest=none -Dsuites="org.apache.comet.exec.CometNativeShuffleSuite" -Pspark-3.5

[mbutrovich]

Thanks @0lai0! Starting CI and will review today.

[mbutrovich]

Looks right to me, thanks for the quick fix @0lai0! Approved pending CI.

[0lai0]

Thanks @mbutrovich for reviews. : )