qwen4exp: port selected upstream performance paths

[edwinbrowwn]

What changed

• Use the KV sequence-position index for n-gram history lookups.
• Reduce Qwen4Exp indexer scoring copies.
• Add opt-in QSA gather for supported flash-attention decode, preserving selection budgets, masked padding, multi-stream bias strides and cache-layout guards.
• Add opt-in Qwen4Exp PLE explicit reads with lazy-mmap/resident fallback, file-span validation and safe reader ownership.
• Port upstream models: fix GDN normalization from max to rsqrt ggml-org/llama.cpp#28068 GDN normalization correctness fix across affected architectures; RWKV normalization is unchanged.
• Keep existing rocPRIM TOP_K unchanged: the temporary radix experiment did not establish a valid kernel-level comparison.

Validation

• CPU/HIP Release builds and argument-parser tests pass.
• Qwen4Exp replay passes with gather enabled/disabled, including individual V620 GPUs and TP4 F16/Q8 configurations. TP roundtrip remains skipped by the harness.
• Independent GDN scalar-reference tests cover widths 4/32/128, epsilons 1e-2/1e-6, ordinary/tiny/epsilon-scale/zero rows on CPU and HIP devices.
• PLE row tests cover F32/F16/Q8_0/Q4_0, shuffled/duplicate rows and multiple workers; F32 bounds/truncation tests pass. Temporary paths are cleaned on early returns and exceptions.
• A compatible synthetic Qwen4Exp GGUF fixture, not a production model, matched direct-vs-lazy-mmap logits exactly. Resident fallback with load-mode none also matched exactly.
• DFlash controller, sidecar-cap, speculative-content and sidecar-artifact tests pass. Server/sidecar builds pass.
• git diff --check passes. Local logs: /home/edwin/.ralph/merge-review-.log and final-.log.

Limits and merge gate

QSA gather and PLE direct reads remain opt-in. No throughput or memory improvement is claimed. Production Qwen4Exp PLE benchmarking remains blocked by available compatible model artifacts. MTP gather, QSA input sharing, PLE graph splitting and incremental block-key caching are separate work.

The prior CI run (34130448354) has ARM64 failures in files unchanged from fork/master: llama-bench.cpp unused repetitions parameter with -Werror, and Windows ARM64 test-recurrent-state-rollback missing llm_arch_supports_rs_rollback at link time. New-head CI must be reviewed before merge; local CPU/HIP success is not a claim that cross-platform CI is green.