Add SD-loaded data and pretrained model support

[RealBhupesh]

Summary

This PR adds support for loading training data from SD-backed CSV files and for running prediction directly from a pretrained model without calling train() on-device.

Changes

• add PLS::setModel(const MatrixXf& B, const MatrixXf& meanX, const MatrixXf& meanY)
• add reusable CSV/model loading helpers in PLSduinoIO.h and PLSduinoIO.cpp
• validate pretrained model dimensions before accepting B, meanX, and meanY
• make predict() fail cleanly when the model is missing or the input feature count is wrong
• reject empty training data in train()
• update the SD-card training example to use shared CSV loading helpers and MatrixXf consistently
• add a new example for loading B.csv, meanX.csv, meanY.csv, and predicting directly from SD
• update the README to document the new workflows and required CSV shapes

Notes

• prediction-only mode requires B, meanX, and meanY; loading B alone is not sufficient because prediction applies centering and de-centering
• SD initialization remains sketch-owned so users can keep using SD, SD_MMC, or another filesystem backend

Testing

• verified the implementation statically in the repo
• could not run Arduino example builds in this environment because arduino-cli is not installed