Default TensorRT transformer trunks to NHWC

[zsqdx]

Summary

• Default TensorRT transformer trunks to NHWC.
• Keep trtTransformerNHWC=false as an explicit per-machine fallback.
• Normalize non-transformer models to NCHW so convnets do not create redundant plan/timing cache entries.
• Document the default and override in gtp_example.cfg.

Motivation

The old NCHW default was based on earlier TensorRT results. Current results point the other way: my RTX PRO 6000 Blackwell test with TensorRT 10.16.1 showed a large NHWC win, and follow-up testing by lightvector on a weak laptop GPU and an A5000 found NHWC equal or faster on both TensorRT 10.9 and 10.16.

Official katago benchmark results for b11c768h12nbt3tflrs-fson-silu, FP16, fixed symmetry and seed:

Backend/layout	Recommended threads	visits/s	nnEvals/s
CUDA	64	2814.27	2716.93
TensorRT NCHW	40	1625.87	1531.59
TensorRT NHWC	32	2836.89	2612.18

Equal-thread NHWC gains over NCHW were +29.1% at 5 threads, +55.8% at 16, +81.2% at 32, and +78.6% at 64.

The resolved layout is already part of the timing-cache hash and the onnx / onnxnh plan-cache tag, so the two engine layouts cannot collide.

Validation

• Fresh TensorRT build at this PR's exact commit against TensorRT 10.16.1 / CUDA 13.
• With no trtTransformerNHWC key, runtime logged transformerNHWC=true: 2852.22 visits/s, 2622.04 nnEvals/s at 32 threads.
• With explicit trtTransformerNHWC=false, runtime logged transformerNHWC=false: 1578.90 visits/s, 1453.44 nnEvals/s at 32 threads.
• OpenCL full build, katago runtests, and katago runoutputtests.