[codex] refactor shared biased-linear kernel

[adrienlacombe]

Summary

• add pwm_core::predictor::batched_linear as the shared sequence-row variant of the biased linear kernel
• route prove_block BatchedLinear through the shared helper instead of keeping a local row loop
• add deterministic random cross-check coverage that the flat and batched kernels produce Freivalds-accepted outputs

Root Cause

Issue #182 tracks drift risk from multiple biased-linear implementations. The flat helper was already shared, but the batched block prover still carried its own row-wise expansion. This keeps the batched path behind the same core kernel boundary as the reference and flat prover paths.

Closes #182.

Validation

• cargo test -p pwm-core linear_kernels_match_freivalds_relation
• cargo test -p pwm-verifier --test block accept_multiposition_attention_with_projection
• cargo test --workspace
• cargo clippy --workspace --all-targets -- -D warnings