Carry per-group slot mappings into multi-KV forwards

Row: MODEL-MM-deepseek-v4-deepseek-v4-for-causal-lm

DeepSeek V4 Vision W4 found that StepInputs::slot_mapping[g] is computed for every KV group, but the runner forwards only the full-attention group's mapping through CommonAttentionMetadata. MultiKvCacheIndex carries per-group block tables but no parallel per-group slot mappings.

That makes a production multi-cache model validate the real cache groups and then either invent identity slots or ignore non-primary cache writes. W4 needs the actual SWA, compressed-latent, attention-compressor, indexer-key, and indexer-compressor slots.

Fix in the same flow: add a borrowed per-group slot-mapping channel to the shared multi-KV index, populate it from runner-owned step inputs for the forward lifetime, and mutation-prove a non-identity mapping through the registered DeepSeek V4 path. The pull request for #2411 closes this issue too.

Trace StepInputs::slot_mapping through CommonAttentionMetadata and MultiKvCacheIndex, starting with the shared multi-KV forward flow. Determine how runner-owned mappings live for the forward and how the registered DeepSeek V4 path consumes them. Done means each KV group receives its actual mapping, and a non-identity mapping is mutation-proven through that path.