feat: [ENG-2700] dedupe duplicate models from openai-compatible upstream (#652)

* feat: [ENG-2700] dedupe duplicate models from openai-compatible upstream

NVIDIA NIM's /v1/models endpoint returns byte-identical duplicate entries
(openai/gpt-oss-120b, openai/gpt-oss-20b, nvidia/nemotron-3-super-120b-a12b),
which surfaced as duplicate rows in the /model TUI dialog when users
configured openai-compatible against NIM. Dedupe by id (first occurrence
wins) inside OpenAICompatibleModelFetcher so any aggregator/proxy that
violates the OpenAI spec doesn't leak duplicates into the dialog.

* feat: [ENG-2700] skip upstream model entries with no id or name

Defensive guard prompted by PR #652 review: previously an upstream entry
with neither `id` nor `name` would collapse to a single record with empty
id/name (broken row in /model). Skip such entries entirely.