feat(rocke): GDN decode kernel and dispatch for gfx950

[AviralGoelAMD]

ISSUE ID : AICK-2228

Adds the GDN (Gated Delta Network) decode kernel for gfx950 — the single-token step run once per generated token — plus dispatch wiring, tests and docs. 21 files, +3282.

Why: softmax decode re-reads the whole KV cache, so cost grows with sequence length. GDN keeps a fixed-size state per head: decay it, add a rank-1 update from the new key/value, read the output. Memory per sequence is constant.

Work split	one workgroup per (sequence, value head, v-sub-block)
State tile	register-resident — no LDS, no barriers
Cross-lane	xor butterfly only: quad_perm at offsets 1–2, ds_swizzle above
State storage	paged pool via read_indices / write_indices; negative index = idle lane
Pool addressing	64-bit byte offset, so a deep pool does not wrap signed 32-bit
Tile knobs	batch-banded table from a correctness-gated sweep; tune.py regenerates it
Support check	dispatch calls the kernel's own is_valid_spec — one rule, not two copies that drift

Stack: PR 3 of 4 — quad_perm (ROCm#12070) → dispatch.core hoist (ROCm#12090) → this → GDN prefill. Needs both, so it sits on an integration of develop + those two and builds standalone.

Verification — head 4abab8e95d1

check	result
CPU suites	60 passed, 0 skipped
Golden IR, llvm20/22/23	21/21 SHAs byte-identical
AICK-2228 criteria on gfx950	6/6 PASS
Host-validation lanes	23 passed, 0 skipped

Output and state checked against an independent fp32 oracle inside TOL = 1e-2; idle lanes bit-identical; deep-pool lane allocated past 2³². Performance numbers live in the protected results page.

Caveat, scoping all of the above: rocke_engine cannot be built on that node, so the run exercised Python-lowered code objects. The C++ path is covered by the byte-identity gate, not by execution.

For reviewers

Every gate was made to fail, not just pass — the numeric gate rejects a wrong output and a wrong state; the tuner gate rejects a corrupted reference; the new untouched-page check rejects a spilled write while that tile's written-pages term stays inside tolerance, which attributes the rejection to the new check.

The test fixture changed. The pool used to be exactly batch deep with read_indices == write_indices, so every page was written every launch and a correct value written to the wrong slot was invisible. Same injected fault, two geometries:

pool	a write that aliases its read slot
old, fitted	1.949e-03 — identical to a clean run, missed
new, disjoint slots	9.780e-01 — rejected

Pool now defaults to 2 * batch + 1 with disjoint slots; check() and the tuner compare untouched pages against their pre-launch contents. make_inputs still exposes pool_depth / disjoint_writes, but the detecting shape is the default.

[AviralGoelAMD]

Superseded by the upstream PR: ROCm#12172 (same content, rebased onto develop so the diff is the 21 decode files only). Closing this working copy.