feat(onnx): add DirectML/MIGraphX/CoreML execution provider support

[noctrex]

Extend colgrep's ONNX Runtime integration beyond CUDA so the same code path can drive non-NVIDIA GPUs via DirectML (Windows), AMD GPUs via MIGraphX, and Apple Silicon via CoreML, selected by cargo feature. The active backend is now also surfaced in the user-facing Model line.

next-plaid-onnx/src/lib.rs

• Add impl Display for ExecutionProvider with short canonical labels (CPU, CUDA, DirectML, CoreML, MIGraphX, TensorRT, auto) matching the tokens already used in onnx_runtime.rs download messages.
• Add test_execution_provider_display alongside the existing EP tests.

colgrep/src/index/mod.rs

• New cfg-gated IndexBuilder branch for directml/migraphx/coreml features that initializes the runtime and picks the matching ExecutionProvider. Falls through to the existing CPU branch when none of those features are enabled.
• Searcher::Auto (two sites) extends its fallback chain past CoreML to also consider DirectML and MIGraphX before Cpu.
• The Model: eprintln! now appends the selected backend in parentheses, e.g. "Model: lightonai/LateOn-Code (DirectML)". Provider is resolved at runtime so a CUDA build that falls back to CPU honestly prints CPU.

colgrep/src/onnx_runtime.rs

• Drop the USE_GPU bool constant in favor of per-feature eprintln!s so each provider gets an accurate label at download time.
• Route DirectML into the existing "gpu" cache subdir (shared with CUDA) and keep CPU in "cpu".
• Add a NuGet download path for Microsoft.ML.OnnxRuntime.DirectML on win-x64, since the GitHub release artifacts do not ship DirectML. Other platforms/configs keep using GitHub releases unchanged.

Note: MIGraphX and CoreML are wired into provider selection but their runtime download still falls through to the CPU package; users on those platforms are expected to supply the matching ORT build out-of-band for now. DirectML is the only new provider with full auto-download support.

Verified:

• cargo test -p next-plaid-onnx execution_provider (6 passed)
• cargo build --release -p colgrep --features directml on Windows x86_64
• cargo build --release -p colgrep (default features)
• Runtime: colgrep prints "Model: lightonai/LateOn-Code (DirectML)" with --features directml

Disclaimer: This commit was authored with assistance from an LLM.

[aussetg]

Did you bench each backend ? In my own tests, Migraphx is slow ( see #122 ). But I can't test the rest.

[noctrex]

@aussetg I am on Windows 11, with a AMD Ryzen 5800X3D CPU and a AMD RX 7900XTX GPU.
I just tried to index this very own codebase using the model LateOn-Code, using the DirectML backend to utilize the GPU.
Only this backend is available in Windows for AMD.

DirectML: 70,87 seconds
CPU: 643,25 seconds

Without this PR, it would not use DirectML at all and was always using only CPU.

[raphaelsty]

Amazing MR thank you 😊