perf(qwen3.6): default the merged gdn_ar_fast_kernel ON

[jaso0n0818]

Summary

SPARKINFER_GDN_FAST (merged in #229) only activates gdn_ar_fast_kernel when the
env var is explicitly set to 1. The eval harness (bench/scripts/evaluate.sh,
evaluate_dual.sh) never sets feature env vars -- it only passes
guard/baseline/difficulty scoring parameters (SPARKINFER_GUARD_*,
SPARKINFER_DIFFICULTY_*, SPARKINFER_EVAL_MODE, etc.) -- so the merged kernel has
never actually run during scoring. The naive gdn_ar_kernel
(<<<v_heads=32, head_dim>>>, ~81% of the SMs idle every token, two serial
head_dim-long dependent-load loops, as described in #229's own summary) has stayed
the decode path's largest single non-GEMV cost the whole time.

This flips the default to ON (opt-out via SPARKINFER_GDN_FAST=0), matching
every other feature flag in this runtime (use_pq, use_llama, use_q6mmvq,
use_fnq are all default-on / opt-out-via-0). No kernel code changes --
gdn_ar_fast_kernel is #229's kernel, unmodified; this is a one-line default flip
plus a comment explaining why.

Scope: kernels/csrc/cuda/fused/qwen36.cu, single function
(launch_qwen36_gdn_ar), single line changed.

Proof of speedup

• Tested on RTX 5090 (sm_120)

Decode tok/s (bench/scripts/bench.sh, Qwen3.6-35B-A3B-UD-Q4_K_M.gguf, bs=1, same box):

context	before (main default)	after (this PR)	delta
128	171.90	224.47	+30.6%
512	170.46	222.07	+30.3%

=== before (main, SPARKINFER_GDN_FAST unset -> naive gdn_ar_kernel) ===
model        : Qwen3.5/Qwen3.6-35B-A3B hybrid  (40 layers, 256 experts top-8)
VRAM used    : 25.3 GB
decode tg    : 171.90 tok/s  (5.8 ms/token, n=128, ctx=0, bs=1)
decode tg    : 170.46 tok/s  (5.9 ms/token, n=512, ctx=0, bs=1)

=== after (this PR, default now runs gdn_ar_fast_kernel) ===
model        : Qwen3.5/Qwen3.6-35B-A3B hybrid  (40 layers, 256 experts top-8)
VRAM used    : 25.3 GB
decode tg    : 224.47 tok/s  (4.5 ms/token, n=128, ctx=0, bs=1)
decode tg    : 222.07 tok/s  (4.5 ms/token, n=512, ctx=0, bs=1)

Correctness (teacher-forced, same held-out prompt, vs main default):

• generate(): token-for-token identical across 48 generated tokens.
• score: PPL 2.76957 -> 2.76956, ARGMATCH 73/101 -> 71/101 (0.7228 -> 0.7030,
  both far above bar). The tiny ARGMATCH delta is the same kind of fp32
  reduction-order noise any split/tiling change in this codebase produces (feat(qwen36): add optimized Gated-DeltaNet AR state update kernel wit… #229's
  own PR notes this is byte-exact math, different order); PPL is unchanged to 5
  significant figures.
• ctest --test-dir build: 7/7 pass.

Why this wasn't caught by #229's own eval

#229 was rebased on top of #230 (shared-expert fix, 23->167 tok/s) before its final
eval run, and #230's fix alone is a 7x win -- enough to land eval:XL even with
gdn_ar_fast_kernel sitting inactive behind its opt-in flag the whole time. The
gap only shows up when you diff default-vs-explicit-flag on the current main,
which is what this PR does.

[ai-hpc]

⏳ Needs a benchmark to be evaluated

You ticked Tested on RTX 5090 but the decode before → after tok/s table is still empty / placeholder (or shows no gain). The on-device eval won't run until it shows a real improvement.

Fill it from the end-to-end decode bench (not an isolated-kernel microbench):

bench/scripts/bench.sh --download            # baseline (before)
bench/scripts/bench.sh --download            # your branch (after)

Then the bot greenlights it on the next poll and evaluates it on an RTX 5090.