fix: support OSS model discovery in Pi

[dairyqiu]

Problem

Pi only configured Claude, OpenAI Responses, and Gemini providers. Workspace model discovery already identifies Kimi, GLM, and DeepSeek services, but Pi did not consume that OSS family or have a provider that could route it.

OpenCode works because it already maps OSS models to the AI Gateway MLflow chat-completions route. This is a Pi integration gap, not a Databricks AI Gateway limitation.

Fix

• Discover and persist OSS models when configuring Pi.
• Add a managed databricks-oss provider using Pi openai-completions against /ai-gateway/mlflow/v1.
• Keep per-request credentials through !ug auth-token.
• Disable unsupported store and strict-mode fields for the MLflow route.
• Apply known per-model context and output limits, including GLM output caps.
• Support managed allowlists, default selection, availability checks, and explicit model selectors for OSS models.
• Remove the stale databricks-kimi provider written by the earlier implementation.
• Extend the live Pi e2e harness to discover, require, configure, and invoke every OSS model exposed by the test workspace.

This revisits #227 on current main. The earlier PR stopped at config-shape validation; this version validates live discovery and inference with the installed Pi client.

Verification

• uv run ruff check .
• uv run ruff format --check src/ tests/
• Final focused suite: 387 passed.
• Final full suite: 2309 passed, 42 skipped. The only failure is the existing installed Claude Code user-agent canary rejecting its synthetic model before reaching the capture server.
• Isolated Pi 0.84.3 model listing: live workspace discovery returned seven OSS models, all shown under databricks-oss while the developer state and real Pi config remained unchanged.
• Isolated live OSS inference: all seven discovered models passed, covering three DeepSeek, three GLM, and Kimi K3 models through /ai-gateway/mlflow/v1/chat/completions.
• The aggregate local Pi e2e run also exercised its existing non-OSS families; its only failures were two deprecated Gemini 2.5 endpoints on the local workspace.

Residual risk

The live checks cover one-turn inference for every discovered OSS model. A multi-turn OSS tool-call loop is not yet covered.

[dairyqiu]

@rohita5l, could you review this PR as the approver? It adds Pi OSS discovery and MLflow chat-completions routing, with live one-turn inference passing for all seven discovered DeepSeek, GLM, and Kimi models. The branch is rebased on current main, and the focused suite and static checks pass.