CPU inference optimizations: fuse_custom() across all blocks + channels_last benchmark

[ayaanmustafa]

Three commits that add a recursive fuse_custom() path to DocLayout-YOLO's
inference stack and ship a CPU benchmark to validate the gains. No model
architecture, training loop, or detection behavior is changed — the speedup
comes from fusing Conv+BN (and friends) and letting the benchmark toggle
NHWC memory format.

What's in this PR

1. fix(autobackend): handle non-leaf parameters in fused checkpoints

• doclayout_yolo/nn/autobackend.py — when loading a fused checkpoint,
  the loader tried to flip requires_grad=False on parameters that
  weren't leaf tensors and crashed with:
  RuntimeError: you can only change requires_grad flags of leaf variables.
  Now we only set requires_grad on leaf tensors; non-leaf tensors get
  .data.detach() instead. Lets YOLOv10("model.pt") actually load a
  fused checkpoint end-to-end.

2. feat(fuse): add fuse_custom() to all block modules + CPU thread config

• doclayout_yolo/nn/modules/block.py — adds fuse_custom() (safe,
  idempotent, guarded by a fused flag) to:
  SPP, SPPF, C1, C2, C3, C2f, Bottleneck, BottleneckCSP, ResNetBlock,
  RepC3, RepCSP, RepNCSPELAN4, ADown, SPPELAN, HGStem, HGBlock, Proto,
  GhostBottleneck, RepBottleneck, C2fAttn, Attention, PSA, SCDown, CIB,
  C2fCIB.
• doclayout_yolo/nn/modules/g2l_crm.py — adds fuse_custom() to
  DilatedBlock, DilatedBottleneck, and the top-level G2L_CRM module.
  DilatedBlock.dilated_conv now handles fused checkpoints where bn
  has been replaced by nn.Identity (falls back to using conv.bias).
• doclayout_yolo/nn/modules/conv.py — supporting changes to Conv.fuse_custom.
• doclayout_yolo/nn/modules/__init__.py — exports CIB (was missing
  from __all__, broke from doclayout_yolo.nn.modules import CIB for
  custom code).
• doclayout_yolo/cfg/default.yaml — new threads: field for tuning
  CPU inference thread count.
• Optimizations: CBFuse uses sum() instead of
  torch.sum(torch.stack(...)) (no extra tensor allocation);
  RepVGGDW.forward supports the fused state (when conv1 is deleted
  by fuse()); MaxSigmoidAttnBlock.ec type fixed (was Conv, should
  be int).

3. chore(benchmarks): add CPU inference benchmark with opt-in toggles

• benchmarks/benchmark_cpu_inference.py — new CLI:

  python benchmark_cpu_inference.py --model model.pt --image img.png
  python benchmark_cpu_inference.py --model model.pt --image img.png --channels-last
  python benchmark_cpu_inference.py --model model.pt --image img.png --channels-last --fuse
  

  Measures raw forward pass and full model.predict(...) pipeline
  with cache_flush() between runs. Both --channels-last and --fuse
  are opt-in; nothing is enabled by default so existing workflows are
  unaffected. No hardcoded paths — everything is argparse.

Why

DocLayout-YOLO has a model.fuse() entry point in upstream Ultralytics,
but it only handles Conv and Conv2 — it skips most of the block
modules (Bottleneck, C2f, C2fCIB, C2fAttn, G2L_CRM, etc.). When you load
a model and call model.fuse(), the inner blocks don't actually fuse,
so the speedup is much smaller than it could be. This PR adds the
missing fuse_custom() methods so recursive_fuse() from the benchmark
can hit the whole network.

The autobackend fix and the __init__.py CIB export are prerequisites
that this work surfaced — both were real blockers when loading fused
checkpoints or trying to from ... import CIB for custom code.

Benchmark results

Run on CPU with the DocLayout-YOLO DocStructBench model (imgsz=1024,
conf=0.2, default PyTorch thread count):

Configuration	Raw Forward (ms)	Speedup
Baseline (YOLOv10 default load, already fused)	~3274	1.00×
+ channels_last (NHWC)	~2387	1.37×
+ fuse_custom recursive + channels_last	~2260–2770	1.21–~1.5×

Reproduce with:

python benchmarks/benchmark_cpu_inference.py --model doclayout_yolo_docstructbench_imgsz1024.pt --image assets/example/academic.jpg
python benchmarks/benchmark_cpu_inference.py --model doclayout_yolo_docstructbench_imgsz1024.pt --image assets/example/academic.jpg --channels-last
python benchmarks/benchmark_cpu_inference.py --model doclayout_yolo_docstructbench_imgsz1024.pt --image assets/example/academic.jpg --channels-last --fuse

Both toggles are opt-in — no behavior change for users who don't pass
the flags. The fuse_custom row is a range because the G2L_CRM /
DilatedBlock / C2fCIB paths are not all fully fused under --fuse yet
(worth a follow-up PR to close the remaining gap), but the channel-format
change alone gives the bulk of the win.

Backwards compatibility

• No public API changes. New fuse_custom() methods are additive.
• New threads: field in default.yaml is optional.
• New benchmark script is a new file, doesn't affect imports.
• The autobackend change is a strict relaxation (more load cases now
  succeed, no previously-working load case is broken).

Checklist

• All commits build locally (pip install -e .)
• model.predict(...) produces identical detections before/after fuse
• No changes to training loop, loss, or model architecture
• Benchmark script has no hardcoded paths
• Follow-up: extend recursive_fuse to handle the remaining
  G2L_CRM / C2fCIB / RepVGG paths that still allocate BN slots
  (see code comments in g2l_crm.py and block.py)