Fix ref batch token normalization

[Connor1996]

What changed

This patch fixes two batching issues in src/tiny_llm_ref/batch.py:

• normalize logits per request by passing axis=-1 to mx.logsumexp(...)
• remove finished requests from every layer cache after decoding, instead of only touching the leaked last batch_cache binding

Why

The main bug was that batched decoding normalized across the wrong dimension, so requests in the same batch interfered with each other and produced broken outputs.

While debugging that path, I also fixed request cleanup so a finished request is removed from all layer caches before its slot is reused.