feat(models): migrate Models APIs to typed NemoClient

[maxdubrinsky]

Summary

• Add typed sync and async Models clients covering the 42 genuine Models CRUD/status/version operations across models, adapters, providers, prompts, deployments, and deployment configs.
• Replace the Stainless-backed packages/models resource implementation while preserving existing URL helpers, polling behavior, pagination, conflict handling, and delete status visibility; sync its vendored package copy without regenerating Stainless.
• Migrate the Models controller, reconcilers, sidecars, model-spec task, workspace cleanup, customization/Data Designer/Automodel/Guardrails/platform-seed/RL/Unsloth consumers, and their tests.
• Preserve the wrapped Stainless SDK retry contract in client_from_platform, including explicit retry headers, 408/409/429, server errors, and reasonable Retry-After values, without changing standalone NemoClient defaults.
• Preserve complete provider metadata during replacement PUTs and omit unrelated partial-update fields so reconciliation cannot clear user-selected routing configuration.

Validation

• 178 client/foundation tests passed.
• 1,205 Models unit tests passed; 1 torch-dependent test skipped.
• 251 downstream consumer tests passed; 1 optional training-image test skipped.
• 49 nemo-unsloth plugin tests passed against typed Models/Files mocks.
• Ruff and formatting passed across all non-generated changed Python files.
• Focused ty, vendored-SDK/CLI sync lints, negative Stainless searches, warning-strict controller checks, and git diff --check passed.
• Multiple independent adversarial review rounds found no remaining blocker or high-severity issue.

Intentional boundaries

• Inference Gateway provider discovery remains on its IGW proxy operation.
• VirtualModel CRUD is owned by AIRCORE-877.
• Files list/download operations and their retry/error handling remain on Files clients or the existing SDK where not part of Models.
• OpenAI proxy, readiness, SSE, and raw-byte operations are not modeled as Models CRUD endpoints.

Follow-up

Directly owned NemoClient transport teardown is tracked separately in AIRCORE-933.

Linear: https://linear.app/nvidia/issue/AIRCORE-876

[github-actions]

Suite	Lines Covered	Line Rate	Branch Rate
Unit Tests	27211/34952	77.8%	61.7%
Integration Tests	15820/33577	47.1%	18.6%