Add Databricks provider

[blairj09]

Adds Databricks as a built-in provider across all three packages, giving Posit Assistant access to models on Databricks Model Serving (foundation models, external models, and custom serving endpoints).

Part of posit-dev/assistant#1110.

ai-provider-bridge

• providers/databricks-provider.ts — model discovery and chat:
  • Discovery: lists chat-capable endpoints from GET /api/2.0/serving-endpoints (READY + llm/v1/chat, including external-model entities). When the workspace has Unity AI Gateway enabled, discovery switches to GET /api/2.0/serving-endpoints:foundation-models, filtered to gateway-v2 chat-capable models.
  • Gateway detection: a probe of /api/ai-gateway/v2/endpoints is cached per workspace host (definitive 200/404/403 results only; transient failures fall back to model serving without poisoning the cache). The cache is shared between discovery and chat so both route consistently, and clears with the model cache.
  • Chat: OpenAI-compatible chat completions through the existing OpenAIClient against {host}/serving-endpoints, with a fetch wrapper that rewrites requests to {host}/ai-gateway/mlflow/v1 when the gateway is available. Credentials are bearer-token apikey credentials (PAT or OAuth access token — the host application decides which) with the workspace host as baseUrl.
• model-capabilities/databricks-helpers.ts — capability inference from the underlying model identity (foundation-model / external-model / endpoint name), reusing the Anthropic and OpenAI helpers after stripping databricks- / system.ai. prefixes. Thinking-effort levels are deliberately dropped for v1; unrecognized models get conservative defaults.
• Registered in PROVIDER_IDS, PROVIDER_MAP (auth provider databricks, apikey), PROVIDER_REGISTRARS, and the public exports. customHeaders are supported on both discovery (additive) and chat paths.

ai-credentials

• CredentialConfig.getDatabricks() + a Databricks branch in shapeCredentials that resolves the workspace host into the credential baseUrl via a new normalizeDatabricksHost helper (accepts scheme-less hosts and trailing slashes).
• The Positron backend reads authentication.databricks.credentials.DATABRICKS_HOST with a process.env.DATABRICKS_HOST fallback for managed environments (e.g. Posit Workbench injecting it into sessions).
• DATABRICKS_TOKEN added to the secret env mappings (store-backend).

ai-config

• databricks added to BUILTIN_PROVIDER_IDS and CLIENT_KIND_VALUES (identity mapping; not offered as a custom-provider type), keeping the cross-package shape guard green.
• The workspace host lives in a dedicated strict databricks: { host } connection section — not baseUrl. Per-model endpoint resolution falls back to the provider baseUrl, so storing the host there routed chat to https://<host>/chat/completions (404) and bypassed the gateway rewrite; the section keeps the host available to the credential shaper without entering endpoint resolution. This mirrors how Snowflake stores host/account.
• The authentication.* host fragment gains a "databricks" read kind (host + customHeaders → the databricks section), and the DATABRICKS_HOST env var overlays the section in the catalog builder.

Testing

• New unit tests: provider discovery/routing/client factory (23 tests), capability inference, credential shaping, Positron backend host resolution, and the host-fragment read kind; existing ai-config provider counts updated (14 → 15).
• npm run build, check-types (including the shape guard), test, lint, and format:check all pass.
• Verified end-to-end in a dev Positron build against a live Databricks workspace: model discovery, and streamed chat through model serving (companion PRs: posit-dev/assistant#1809, Add Databricks authentication provider positron#14898).

Merge ordering

This PR should land before posit-dev/assistant#1809, whose packages/ai-lib submodule pin references this branch.