Add end-to-end guided workflow command for winml config

Summary

Add a single winml flow <model> command that runs the full model preparation pipeline as an interactive, step-by-step guided workflow — completing each stage, showing a summary, then prompting the user to continue or skip.

Motivation

Currently, preparing a model for WinML deployment requires users to know and chain 6–8 separate commands in the correct order (inspect, export, analyze, optimize, quantize, compile, perf, eval). This creates a steep learning curve for new users and forces experienced users to manually orchestrate the pipeline.

A single guided command lowers the barrier to entry, surfaces the recommended workflow explicitly, and eliminates the need to remember inter-command dependencies or flag syntax.

Proposed Solution

Command signature

# HuggingFace model (full pipeline incl. export)
winml flow ProsusAI/finbert

# Local ONNX file (export step skipped; flow adapts to model format)
winml flow ./model.onnx
winml flow ./build/finbert/

Interactive pipeline UX

Each step runs to completion, prints a summary block, then pauses:

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
 Step 2 / 7 — analyze
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
 ✔  Operators:     48 / 48 supported on QNN
 ⚠  Warnings:      2  (see details below)
 ✘  Blocking:      0

 [Enter] Continue to optimize   [s] Skip   [q] Quit

Two input paths

ONNX-aware step suggestions

When the input is an ONNX file, flow inspects the model format and skips or adjusts steps that are not applicable:

• Already-quantized (QDQ) model → skip quantize, warn that re-quantization may degrade accuracy
• Already-compiled EP artifact → skip optimize / compile, jump to perf / eval
• FP16 model → suggest quantize to INT8 for NPU targets

Non-interactive / CI mode

winml flow ProsusAI/finbert --yes          # auto-continue all steps
winml flow model.onnx --skip quantize,eval # skip specific steps
winml flow model.onnx --steps analyze,perf # run only selected steps

Acceptance Criteria

• winml flow <huggingface_model_id> runs the full HF pipeline interactively
• winml flow <path_to_onnx> runs the ONNX pipeline, adapting steps to model format
• Each step prints a structured summary before prompting
• Prompt supports [Enter] (continue), [s] (skip), [q] (quit)
• --yes flag runs all steps non-interactively (CI-friendly)
• --skip <step,...> allows excluding specific steps
• --steps <step,...> allows running only selected steps
• Already-processed ONNX inputs (QDQ, compiled) are detected and steps adjusted with user-visible rationale
• New command module src/winml/modelkit/commands/flow.py following existing LazyGroup conventions
• Unit tests cover step sequencing, skip logic, and ONNX format detection

Alternatives Considered

• winml build extension: build is scoped to HuggingFace export; conflating it with the full pipeline would bloat the command and break its single-responsibility contract.
• Notebook / tutorial: Good for learning but not a repeatable CLI workflow.

Additional Context

Related existing commands that flow would orchestrate:
inspect, export, analyze, optimize, quantize, compile, perf, eval

The command module system uses LazyGroup (see cli.py) — flow.py should follow the same lazy-import pattern.

Multi-surface opportunity: The recommended flow is not exclusive to the CLI. The foundryTK Template UI is another natural surface for exposing this guided pipeline — a visual step-by-step wizard in the Template UI could mirror the same inspect → export → analyze → optimize → quantize → compile → perf → eval progression, making the workflow accessible to users who prefer a graphical experience over the command line.

Start by reading the existing command modules and cli.py, focusing on the LazyGroup conventions and the inspect, export, analyze, optimize, quantize, compile, perf, and eval entry points. Add src/winml/modelkit/commands/flow.py and tests covering step sequencing, skip logic, and ONNX format detection; done means both input paths, interactive prompts, and the --yes, --skip, and --steps modes meet the acceptance criteria.