Adds breadmill's `full` feature (npu + rocm + cuda together) and switches
the release workflow to build with it. All three backends are ort's
load-dynamic (dlopen) mode, so combining them doesn't require the NPU/ROCm/
CUDA toolkits on the build host -- which backend is actually available is
resolved at runtime via ORT_DYLIB_PATH / the dynamic linker, per whichever
backend is selected for that run.
Also fixes a real bug this surfaced: backend selection checked
`config.model.backend == "rocm"` unconditionally in an if/else-if chain, so
an explicit --cuda (or --npu) flag silently lost to an unrelated `backend`
value already sitting in config.toml. CLI flags now always take priority
over config.
Version bump: breadmill 0.2.0 -> 0.2.1.
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.
