[SPARK-49442][SS][FOLLOWUP] Use monotonic clock for Kafka partition cache TTL

[cloud-fan]

What changes were proposed in this pull request?

Followup to #57351.

This PR measures the Kafka partition metadata cache TTL with System.nanoTime instead of
System.currentTimeMillis. It also represents the cached timestamp with Option[Long] rather than
using zero as a sentinel. Debug logging continues to report cache age in milliseconds.

Why are the changes needed?

TTL measurement should use a monotonic time source. Wall-clock adjustments can otherwise make a
cache entry appear older or younger than it is, causing premature refreshes or extending its
lifetime. System.nanoTime measures elapsed time without that sensitivity.

Does this PR introduce any user-facing change?

No. The cache option and its behavior under a stable clock are unchanged.

How was this patch tested?

Existing cache reuse and expiry coverage was run with Java 17:

build/sbt 'sql-kafka-0-10/testOnly org.apache.spark.sql.kafka010.KafkaOffsetReaderSuite'

All 22 tests passed. No new test was added because the change only substitutes the JVM's monotonic
elapsed-time source; reliably simulating host wall-clock changes is not practical in this suite.

Was this patch authored or co-authored using generative AI tooling?

Generated-by: Codex (GPT-5)

[cloud-fan]

cc @vinodkc @HeartSaVioR

[uros-b]

Thank you @cloud-fan!