docs(benchmarks): Gemma3n decode profile, no fusion justified (#329)

[inureyes]

Summary

Profile-first investigation for #329: profile Gemma3n decode, then add a compiled fusion only if the profile justifies one. It does not, so this PR ships the profiling finding as documentation and adds no kernel. A documented "no fusion justified" outcome is the result the issue explicitly calls for.

Findings (Apple M1 Ultra, MLX 0.31.2, mlx-vlm 0.4.4)

• Decode is ~92% GPU-bound. The Rust graph-build step (forward), where an FFI-crossing fusion would land, is ~7% per token and fully hidden behind async_eval. A fusion that only collapses FFI crossings would regress here, the same outcome as MLXCEL_FUSED_QK_NORM on Qwen3 (1-3.4% slower on M1 Ultra).
• Decode tok/s: e2b-4bit 83.1, e4b-4bit 62.9 (text), 58.7 (image). mlxcel already leads mlx-vlm (68.9 / 52.6 / 47.9) by 1.20-1.23x, and mlx-lm cannot load these multimodal checkpoints, so there is no parity gap to chase.
• The worthwhile Gemma3n fusions already shipped in perf(gemma3n): reduce bf16 decode AltUp/MLP graph overhead #60 (fused MLP bridge, compiled gelu_topk/GeGLU, stacked AltUp) and are active on M1 Ultra (use_fused_decode_path() is true on non-NA hardware).
• Pure weight streaming is ~43% of GPU time; the ~57% structure overhead is dominated by command-buffer dispatch gaps, not fusable kernel compute. Raising MLX_MAX_OPS_PER_BUFFER toward 1000 recovers +11-13% with no code (e2b 82.7 -> 93.2, e4b 63.0 -> 70.0). That is a scheduling knob, not a kernel, and must be hardware-gated because M5 regresses with larger buffers, so it belongs in a separate follow-up issue.
• Text vs VLM: the vision tower is a prefill cost only; decode is the same per-layer structure either way.

What changed

• Add docs/benchmark_results/gemma3n-decode-profile.md: full numbers, pipeline split, one-decode-token op histogram, pure-GEMV streaming floor, command-buffer sweep, text-vs-VLM table, and reproduce commands.
• Link it from the "Decode-gap investigations" list in docs/benchmarks.md.

Test plan

• Decode best-of-3 (caffeinate, cool) on gemma3n-e2b/e4b-4bit via mlxcel-bench-decode.
• Pipeline split via MLXCEL_PROFILE_PIPELINE_DETAIL; op histogram via MLXCEL_EXPORT_DECODE_DOT.
• mlx-vlm reference via mlx_vlm.stream_generate (wall-clock).
• Docs only; no code paths changed.

Closes #329