feat(ds4): add indexed layer-major prefill

[davide221]

Summary

• add typed exact, dense, and sparse DeepSeek V4 prefill modes; exact remains the default
• execute optimized prefill layer-major with batched causal compressor updates and device-resident HC/FFN state
• add the DS4 learned indexer QAT/score/mask kernels and the D=512 compact sparse-attention path
• add gfx1151 ROCmFPX MMQ support, shared activation quantization, and grouped output projection without materialized permutations

The optimized path is explicit (--ds4-prefill sparse). It does not add runtime environment-variable switches or change the default prefill policy.

Performance

DeepSeek-V4-Flash-ROCMFP2-STRIX, gfx1151, ROCm 7.2.2, --chunk 2048 --max-ctx 32768, all experts resident, model top-k 6, no draft, and all prefix/prefill caches disabled.

Prompt	Tokens	Auto clock	Pinned 2.9 GHz
2K	2,131	287.0 tok/s warm	315.0 tok/s warm
8K	7,641	258.0 tok/s	274.9 tok/s median
24K	22,841	225.8 tok/s	237.0 tok/s median

Pinned ranges were 314.9–315.2 tok/s at 2K, 274.1–277.8 at 8K, and 236.4–238.6 at 24K. The first cold 2K request was 285.8 tok/s. The clock pin on the test host was:

echo high | sudo tee /sys/class/drm/card2/device/power_dpm_force_performance_level

This is 300+ tok/s for a warm 2K prompt, not a sustained 300+ tok/s claim at longer context.

Correctness boundary

• exact keeps tokenwise scheduling and does not enter the layer-major path.
• Three identical exact requests produced the same token sequence in this branch. The current fix/feat(ds4): ROCmFPX byte-safe decode + fused-decode path + ds4 AR optimizations #503 baseline varied across identical repeated requests, so this PR does not claim equality to one nondeterministic baseline capture.
• Sparse 2K/8K/24K one-token gates produced the same token id and SHA-256 across every run.
• sparse is experimental and may change logits or full generated sequences. It needs task-level quality validation before it can become a default.

Validation

• full ROCm/gfx1151 build
• test_deepseek4_unit
• deterministic exact-mode repeat capture
• sparse 2K/8K/24K sequence and throughput captures
• git diff --check