Optimize predict() performance through improved aggregate() logic

[RestlessTail]

This PR implements the optimized aggregation method discussed in #369, providing performance improvement for predict() while preserving identical outputs.

Summary of changes:

• Replaced the existing aggregation loop in predict() with a vectorized approach that substantially reduces computation time.
• Added a non-regression test under tests/testthat/test-predict.R, using pre-generated reference outputs stored in inst/testdata/predict.block.pls.rda.

The reference data generation script is also available here