perf(dory): cache affine G1 bases for streaming commitment

[mw2000]

process_chunk and process_chunk_onehot convert the same G1 generators from projective to affine form on every call — once per row of the streaming commitment. For a typical proof this happens hundreds of times, each incurring a batch field inversion that produces the same result.

This PR caches the affine G1 bases in a module-level RwLock<Vec<G1Affine>>, initialized once during setup_prover via G1Projective::normalize_batch. Subsequent calls read from the cache with no conversion overhead. The cache is cleared in reset() for test isolation, and affine_g1_bases_or_init() provides lazy initialization as a fallback.

Also removes the unsafe from_raw_parts pointer aliasing that was previously used to access the G1 slice.