feat: add encode time tracking for shuffle operations

[0lai0]

Which issue does this PR close?

Closes #1212
Part of #3996

Rationale for this change

Comet already reports encoding/compression metrics for native shuffle, but JVM columnar shuffle (CometColumnarExchange) either showed 0 ms or lacked a useful task-level distribution in SQL UI. This made it difficult to compare shuffle behavior across spark.comet.shuffle.mode=native and jvm, and reduced observability for JVM shuffle performance tuning.

What changes are included in this PR?

• Aligned the JVM/native shuffle spill contract so encode/compression timing is propagated end-to-end: native spill results are consumed as (written_bytes, checksum, encode_nanos), with corresponding JNI/Java updates.

• Added shared AtomicLong encode-time accumulators in SpillWriter, CometShuffleExternalSorterSync/Async, and CometDiskBlockWriter so encode time is aggregated correctly across batches, spills, and concurrent sorter/writer instances.

• Added getEncodeNanos() to CometShuffleExternalSorter and wired the accumulated value into the encode_time SQLMetric in both CometUnsafeShuffleWriter and CometBypassMergeSortShuffleWriter.

• Ensured JVM columnar shuffle dependencies carry shuffleWriteMetrics, allowing CometShuffleManager to retrieve and pass encode_time into the shuffle writers.

• Added a regression test to verify encode_time exists and is greater than zero for columnar shuffle workloads.

How are these changes tested?