[Feature] Support batched vector index training with sampling ratio

[jerry-024]

Purpose

This PR updates native vector global index building to train through the VectorIndexTrainer / VectorIndexTraining API instead of materializing all training vectors in one Java float[].

Main changes:

• Bump paimon-vector-index-java to 0.2.0-SNAPSHOT for the separated trainer/training API.
• Add train.sample-ratio for vector index training, with default 1.0 so the default path still trains with all non-null vectors.
• Support both index-level and field-level configuration, for example <index-type>.train.sample-ratio and fields.<field-name>.train.sample-ratio; field-level configuration takes precedence.
• When train.sample-ratio is less than 1.0, sample training vectors evenly from the temporary vector file and feed them to the native trainer in batches, while still adding all vectors to the final index.
• Avoid the previous single large Java training array allocation and protect both training and add batches from oversized Java array allocations.
• Log the selected training sample count and large native-memory estimates.

Tests

[leaves12138]

+1