Add batched preprocessing API

[GiraudJules]

Summary

Add a batched preprocessing path that turns multiple images into one contiguous float32 tensor in Rust.

Changes

• add preprocess_batch(...) to the Rust preprocess module
• add PreprocessBatchResult
• expose preprocess_batch_numpy(...) to Python
• support either an NxHxWx3 uint8 NumPy array or a sequence of HxWx3 uint8 arrays
• extend Rust tests and the Python smoke test

Why

This is the next high-value preprocessing step after single-image preprocess and tensor layout helpers. It reduces Python loop overhead in inference pipelines.

Validation

• cargo fmt --check
• cargo test
• cargo check --features python
• ../../.venv/bin/maturin develop --features python
• ../../.venv/bin/python scripts/python_smoke_test.py