[GDN2] Add fused BT=16 inference kernels for GDN2 prefill

[Ghd-2077]

[Ops] Add fused BT=16 inference kernels for GDN-2 prefill

Optimization

This implementation follows the FlashKDA two-kernel inference design for GDN-2:

• K1 (intra-chunk): fuses optional Q/K L2 normalization, gate activation, GDN-2 vector erase/write WY auxiliary construction, and the block-local inverse for BT=16 chunks.
• K2 (inter-chunk): propagates the recurrent state across chunks for each (batch, head, V-tile), writes the output, and optionally stores the final state.
• The intended benefit is fewer intermediate global-memory round-trips than the baseline multi-kernel chunk path, while retaining GDN-2's independent erase gate b, write gate w, and state_v_first support.

Testing

Environment

Item	Value
Git branch	perf/gdn2-fused-infer
Git commit	9079232
Commit subject	test
Working tree dirty	yes
GPU / driver / memory	NVIDIA H100 80GB HBM3, 535.183.06, 81559 MiB
GPU (PyTorch)	NVIDIA H100 80GB HBM3
Python	3.12.13
PyTorch	2.10.0+cu128
CUDA (PyTorch)	12.8
Triton	3.6.0

Performance

Operation: chunk_gdn2_infer
Speedup: baseline median_ms / fused median_ms

B	T	H	D	Fused (ms)	Baseline (ms)	Speedup
1	8192	96	128	2.027	3.719	1.83x
2	16384	16	128	1.875	2.599	1.39x
4	2048	16	128	0.382	1.098	2.87x
4	4096	64	128	2.416	4.771	1.98x
8	1024	8	64	0.212	1.088	5.13x
8	2048	32	256	3.030	5.143	1.70x

Average speedup: 2.48x (arithmetic mean), 2.24x (geometric mean); best case: 5.13x.

Benchmark commands:

python -m benchmarks.ops.run --op chunk_gdn2_infer --modes fwd --base '' --json benchmark_baseline.json
python -m benchmarks.ops.run --op chunk_gdn2_infer --modes fwd --base '' --json benchmark_fused.json

Correctness

PASS: 5 passed, 0 skipped in 4.20s.

Test node: tests/ops/test_gdn2.py::test_chunk_gdn2_fwd_inference

Test	Status	o diff	o ratio	ht diff	ht ratio	Failure detail
test_chunk_gdn2_fwd_inference[small-fp32]	PASSED	0.000142	0.001581	0.000526	0.001094
test_chunk_gdn2_fwd_inference[l2norm-fp32]	PASSED	0.000109	0.001708	0.000681	0.001063
test_chunk_gdn2_fwd_inference[gate-dt-bias-fp32]	PASSED	0.000254	0.001225	0.000615	0.001076
test_chunk_gdn2_fwd_inference[safe-gate-no-bias-bf16]	PASSED	0.000802	0.003675	0.000819	0.002292
test_chunk_gdn2_fwd_inference[state-v-first-v-ne-k-fp16]	PASSED	0.0001	0.000983	0.000171	0.000296

Correctness command:

python -m pytest -vv -s --log-cli-level=INFO tests/ops/test_gdn2.py::test_chunk_gdn2_fwd_inference