fix(model-switch): keep same-endpoint custom providers with different names as separate picker rows

[craigdfrench]

Problem

Multiple custom_providers entries sharing the same base_url (e.g. an aggregating proxy fronting multiple upstream providers at one URL) collapsed into a single picker row under the first entry's name. Users who configured separate entries for distinct sub-providers behind one endpoint saw all their models merged into one row, defeating the purpose of splitting them.

Fix

Extends the group_key in list_authenticated_providers() with a display-name prefix dimension:

group_key = (api_url, credential_identity, api_mode, headers_identity, display_prefix.lower())

The display prefix is the text before an em-dash (—) or space-hyphen-space (-) in the entry name. This means:

• Same-URL entries with different names (e.g. "My Proxy / cog-proxy", "My Proxy / nvidia") now get separate picker rows — each showing only its own model list.
• Same-URL entries sharing a common prefix (e.g. "Ollama — GLM 5.1", "Ollama — Qwen 3") still collapse into one "Ollama" row — preserving the existing per-model-suffix grouping behavior.
• Same-host entries with different key_env or api_mode stay separate (credential_identity / api_mode preserved — no regression).

Why this is a bug fix, not a feature

The current grouping logic assumes all entries sharing an endpoint are duplicates of the same provider. This is wrong for aggregating proxies — users intentionally configure multiple entries pointing at the same URL to represent distinct upstream providers, each with their own model subset. The picker silently merging them destroys that configuration.

Tests

• test_shared_url_different_display_names_are_separate_rows — three entries sharing base_url+api_key+api_mode but different names → three rows.
• test_shared_url_per_model_suffix_still_collapses — "Ollama — A"/"Ollama — B" still collapse into one "Ollama" row.
• Fixed 3 pre-existing tests that were hitting a live local Ollama during model discovery (added fetch_api_models mock).

Usage

Set discover_models: false and an explicit models list on each sub-provider entry:

custom_providers:
  - name: My Proxy / sub-provider-1
    base_url: http://my-proxy.example.com/v1
    api_key: no-key-needed
    api_mode: chat_completions
    discover_models: false
    models:
      model-1a: {}
      model-1b: {}
  - name: My Proxy / sub-provider-2
    base_url: http://my-proxy.example.com/v1
    api_key: no-key-needed
    api_mode: chat_completions
    discover_models: false
    models:
      model-2a: {}

Each entry appears as its own row in the /model picker.

Supersedes the grouping fix portion of #28218.

[teknium1]

Merged via PR #67925 — your commit was cherry-picked onto current main with your authorship preserved in git log. This was exactly the right shape: extending the existing identity tuple instead of replacing it kept the credential/mode/headers safeguards intact, and the suffix-fold tests made the behavioral contract explicit. One small follow-up on our side: your test-mock additions for the three Ollama-flaky tests were dropped in favor of #62421's probe_custom_providers=False (merged as #67921 just before this).

#67751 (excluded_providers) is next in the review queue now that this base has landed. Thanks!