Fix ROCm backend (target MIGraphX EP) and add CUDA support

ROCm silently fell back to CPU: the code targeted ONNX Runtime's classic
ROCMExecutionProvider, but distro ROCm-enabled ONNX Runtime builds (e.g.
Arch's onnxruntime-rocm) are commonly compiled with --use_migraphx instead,
and registration failures were invisible since breadmill never installed a
tracing subscriber. Switches the rocm feature to target MIGraphX, adds a
default tracing subscriber so EP registration success/failure is always
visible, and fixes a real crash where MIGraphX's output sequence padding
could index the attention mask out of bounds during mean-pooling.

Also adds a CUDA backend (--cuda / backend = "cuda") mirroring the same
ort execution-provider pattern, for NVIDIA hardware.

Version bump: 0.1.0 -> 0.2.0.

DESIGN.md

@@ -80,3 +80,7 @@ Start from `breadbox/breadbox/src/main.rs`. **Reuse verbatim:** the gtk4-layer-s
 - nomic prefixes + mean-pool + normalize must match between index and query or recall collapses.
 - `ort` linking: prefer the crate's downloaded/bundled ONNX Runtime to avoid version skew with Arch's `onnxruntime`.
 - Office formats (docx/odt) are best-effort in v1; md/txt/org/pdf are the reliable path.
+- GPU EPs (ROCm/CUDA) fail to register silently at the ONNX Runtime level and fall back to CPU — always check
+  startup logs for `Successfully registered` before trusting a GPU build is actually accelerating. See
+  [README: GPU backend notes](README.md#gpu-backend-notes) for the MIGraphX-vs-ROCMExecutionProvider distinction
+  and the per-shape JIT-compile-and-cache behavior that matters for interactive query latency.