[SPARK-59052][ML] Reduce LinearSVCModel transform closure size and improve prediction performance

[zhengruifeng]

What changes were proposed in this pull request?

This PR adds protected column-expression hooks to ClassificationModel for producing raw
predictions, converting raw predictions to labels, and predicting labels directly from features.
ClassificationModel.transform uses these hooks while retaining the existing UDF-based defaults.

LinearSVCModel overrides the hooks to snapshot only coefficients, intercept, and threshold in
its UDF closures. Raw-prediction-to-label conversion uses Spark column expressions, and the
model-capturing margin function field is removed.

Why are the changes needed?

The previous transform implementation built UDFs from bound model methods. This serialized the
entire model into each closure and performed parameter lookup for every input row. The hooks let
classification models provide compact closures or native Spark expressions without changing
prediction behavior.

An in-process comparison used 4,096-feature models for closure serialization and 3,000,000 cached
rows with 32 features for 20 alternating execution measurements:

Transform path	Previous closure	New closure	Previous median	New median
Raw prediction	38,268 B	34,627 B (-9.5%)	553.5 ms	527.1 ms (-4.8%)
Prediction only	38,206 B	34,580 B (-9.5%)	426.3 ms	229.2 ms (-46.2%)
Raw + prediction	76,552 B (2 UDFs)	34,627 B (1 UDF) (-54.8%)	818.8 ms	586.7 ms (-28.3%)

The temporary benchmark probe was removed after measurement.

Does this PR introduce any user-facing change?

No.

How was this patch tested?

• build/sbt mllib/scalastyle
• build/sbt 'mllib/testOnly org.apache.spark.ml.classification.LinearSVCSuite'
  (16 tests passed; 1 pre-existing test ignored)

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

Generated-by: Codex (GPT-5)