gpu-gates: q8_2 quantization silently drops a partial block #175

The unchanged loop in ik_q8_2::quantize walks only x.as_chunks::<32>().0. Passing 33 finite values returns 32 codes and one scale without an error. gate_deepseek41_step::v_q8_2 then indexes q[c] for every input value; the Q8_0 project functions also ignore the partial block. The whole-block check lives in reconstruct, so direct callers bypass it. Add that named assertion to quantize and pin rejection of a 33-value input.

Evidence (at the base of #174):

• crates/gpu-gates/src/ik_q8_2.rs:21
• crates/gpu-gates/src/bin/gate_deepseek41_step.rs:3419

_{Filed by Shingler, the review bot, while reviewing #174; the defect is outside that pull request.}