Fix DeepSeek-V4 Native history across prefill chunks

[remer]

Closes #29

Summary

• restore each DeepSeek-V4 Native compressor's packed prefix before appending to the shared BF16 history scratch
• exercise Native FP4 pack/unpack at a high destination-row offset
• compare Native learned-index scores bit-for-bit with the BF16 scorer using the same expanded codes

Root cause

The request-scoped Native history allocation is shared across layers and across the main and learned-index compressors. It therefore cannot preserve one compressor's prefix across intervening owners or later prefill chunks. The previous active_prefill_chunk_idx == 0 guard skipped the required restore after chunk 0, allowing learned-index selection to consume mixed history.

The packed Native cache remains the persistent source of truth. This change restores its prefix whenever first_group > 0 before appending new rows.

Validation

• exact 59,168-token all-Native retrieval test: 4/10 before, 10/10 after
• 103,349-token three-chunk retrieval test: 2/2
• high-offset Native FP4 pack/unpack: bit-exact at destination row 14,792
• Native learned-index scorer versus BF16 scorer over identical expanded codes: bit-exact
• git diff --check

The source-build compatibility change used on the validation host is intentionally excluded from this PR.