Run a denoiser of any input size on TensorRT and CUDA

Role::WholeDenoiser is the denoise network exported with any height and
width, for a whole frame instead of 1408 tiles whose borders are thrown
away. It is served only where a new size costs nothing: the CUDA
provider, and TensorRT through an optimisation profile from 256 to
4608 x 6656, tuned for the 6D's frame with Best's border. Everywhere
else whole_frame_limit() says None and the fixed tiles run.

ort's TensorRT builder has no profile options, so the engine registers
through the runtime's V2 options with the names 1.30 reads
(trt_profile_{min,opt,max}_shapes). Without a profile a dynamic input
compiled an engine per size at run time, 156 s on the first frame. The
engine lives in its own directory per model: ORT's cache key leaves the
shape out, and the fixed 1408 export and its any-size sibling are the
same graph.

core/dr-inference-engine/src/engines.rs

@@ -68,6 +68,16 @@ pub fn openvino_dir(cfg: &Config, bytes: &[u8], fp16: bool) -> PathBuf {
     )
 }
 
+/// Where TensorRT keeps the engine for a whole-frame model. Its own
+/// directory per model: ONNX Runtime's engine cache key leaves the input
+/// shape out, and served one export's engine to another of the same graph
+/// with a different shape when the denoiser was first cut into pieces
+/// (2026-10-04) — the fixed 1408² denoiser and its any-size sibling are
+/// exactly that pair.
+pub fn tensorrt_whole_dir(cfg: &Config, bytes: &[u8]) -> PathBuf {
+    model_dir(cfg, "tensorrt-whole", bytes)
+}
+
 /// `<cache>/<provider>/<runtime version>/<hash of the bytes>`: one per
 /// model, and one per runtime version, which wrote it.
 fn model_dir(cfg: &Config, provider: &str, bytes: &[u8]) -> PathBuf {