feat: add TopK, Gather, GatherElements, Expand, Tile, and Mod operators

[vlordier]

Summary

Adds six missing ncnn operators required for YOLOv10/YOLO26-style model deployment, along with pnnx lowering passes and full test coverage. Consolidates and supersedes #6558 and #6668.

Operators Implemented

Operator	Layer	pnnx pass	Tests
TopK	src/layer/topk.cpp	pass_ncnn/TopK.cpp	tests/test_topk.cpp
Gather	src/layer/gather.cpp	pass_ncnn/torch_gather.cpp	tests/test_gather.cpp
GatherElements	src/layer/gatherelements.cpp	pass_ncnn/gatherelements.cpp	tests/test_gatherelements.cpp
Expand	src/layer/expand.cpp	pass_ncnn/expand.cpp	tests/test_expand.cpp
Tile (ONNX)	src/layer/tile.cpp (extended)	pass_ncnn/tile.cpp	existing
Mod	src/layer/mod.cpp	pass_ncnn/mod.cpp	tests/test_mod.cpp

Review Issues Addressed

All Copilot review comments from #6558, #6668, and the initial #6669 review have been resolved:

• int32/int64 index dispatch: READ_IDX macro in Gather and GatherElements branches on idx_elemsize (4 vs 8)
• Output rank preservation: top_blob allocated with overload matching index_blob.dims (1D/2D/3D)
• Correct axis ordering: Gather/GatherElements use PyTorch-style axis (0=outermost); TopK uses ncnn-internal axis (0=w=innermost); pnnx TopK pass now converts new_axis = (ncnn_ndim-1) - new_axis
• TopK int32 indices: TopK outputs int32 indices (not float), enabling correct downstream Gather indexing
• TopK pnnx axis fix: pnnx TopK pass converts PyTorch dim to ncnn-internal axis ordering
• pnnx TopK writes k: op->params["3"] = k_val now correctly propagated
• expand.cpp int64 shape support: shape blob read via shape_elemsize (4 vs 8), validation rejects invalid broadcasts
• expand.cpp total() padding bug: replaced with explicit w*h*c loops; no NEON intrinsics
• torch_gather.cpp batch_index: applies axis > batch_index ? axis-1 : axis remapping
• onnxruntime opt-in: -DPNNX_DISABLE_ONNXRUNTIME=ON cmake option; not forced-disabled
• Test coverage: value-checking tests for all six operators; sorted=0 paths covered in test_topk

End-to-End Verification

pnnx converts a torch.topk + torch.gather TorchScript model ([1,8]→top-5) and ncnn inference matches PyTorch reference:

topk values: [0.9000, 0.8000, 0.7000, 0.5000, 0.4000]
gathered:    [0.9000, 0.8000, 0.7000, 0.5000, 0.4000]  ✓

Test Results

167/167 tests passed, 0 failed

[vlordier]

All Copilot review comments have been addressed in the latest commits. Summary of fixes:

Comment	Status
Expand not in ncnn_add_layer	✅ Fixed — registered at line 108 of src/CMakeLists.txt
Gather/GatherElements output always 3D	✅ Fixed — top_blob.create() overload matches index_blob.dims
Index tensor read as const int* ignoring int64	✅ Fixed — READ_IDX macro branches on idx_elemsize (4 vs 8)
GatherElements axis dimension order wrong	✅ Fixed — PyTorch-style axis ordering (axis=0=outermost)
expand.cpp shape_blob read as const int* (no int64)	✅ Fixed — reads via shape_elemsize check
expand.cpp no broadcast validation	✅ Fixed — validates each dimension, rejects invalid shapes
torch_topk.cpp header counts 12 7	✅ Fixed — now 7 7 matching actual operators
ir.cpp TopK emits numeric param ids (0=...)	✅ Fixed — emits named args (k=, axis=, largest=, sorted=)
tensor_to.cpp type_from=0 unconditional	✅ Already correct — looks up pnnx_to_ncnn_cast_type[in_pnnx_type]
pnnx CMakeLists forces onnxruntime_FOUND=FALSE	✅ Fixed — PNNX_DISABLE_ONNXRUNTIME opt-in option (default OFF)
No tests for Gather/GatherElements/Expand	✅ Fixed — test_gather.cpp, test_gatherelements.cpp, test_expand.cpp, test_mod.cpp added
TopK pass never writes k	✅ Fixed — op->params["3"] = k_val
torch_gather.cpp ignores batch_index	✅ Fixed — applies axis > batch_index ? axis-1 : axis
New (this session): TopK pnnx axis wrong for ncnn-internal ordering	✅ Fixed — new_axis = (ncnn_ndim-1) - new_axis
New (this session): TopK outputs float indices, breaks Gather	✅ Fixed — TopK now outputs int32 indices

End-to-end verified: torch.topk + torch.gather TorchScript model converted by pnnx and executed by ncnn produces [0.9, 0.8, 0.7, 0.5, 0.4] matching PyTorch reference exactly.

All 167 ncnn tests pass.

[vlordier]

End-to-end YOLOv10n conversion verified ✅

Converted YOLOv10n (ultralytics 8.4.37) to ncnn using pnnx with the operators from this PR:

python3 -c "
from ultralytics import YOLO
YOLO('yolov10n.pt').export(format='torchscript', imgsz=320)
"
pnnx yolov10n.torchscript inputshape=[1,3,320,320]f32

Result: 268 layers, 0 ignored ops, with TopK and Gather correctly lowered:

TopK    topk_162    1 2 296 297 298  0=-1 1=1 2=1 3=300
Gather  gather_165  2 1 294 302 303  0=0
TopK    topk_163    1 2 304 305 306  0=-1 1=1 2=1 3=300
Gather  gather_166  2 1 292 317 318  0=0

Before this PR, both torch.topk and torch.gather appeared as ignore in the conversion output. Now they produce correct ncnn layers with proper axis and k parameters.

Environment: macOS (Apple M4 Pro), torch 2.8.0, ultralytics 8.4.37

[vlordier]

YOLO26n conversion verified ✅

Also converted YOLO26n (the original motivation for this PR) with identical success:

python3 -c "
from ultralytics import YOLO
YOLO('yolo26n.pt').export(format='torchscript', imgsz=320)
"
pnnx yolo26n.torchscript inputshape=[1,3,320,320]f32

Result: 337 layers, 0 ignored ops, TopK and Gather fully lowered:

TopK    topk_205    1 2 380 381 382  0=-1 1=1 2=1 3=300
Gather  gather_208  2 1 378 386 387  0=0
TopK    topk_206    1 2 388 389 390  0=-1 1=1 2=1 3=300
Gather  gather_209  2 1 376 401 402  0=0

Both YOLOv10n (268 layers) and YOLO26n (337 layers) convert cleanly end-to-end.

Environment: macOS (Apple M4 Pro), torch 2.8.0, ultralytics 8.4.37