feat(serve): remove LLM endpoints over the management API

[ArtemisMucaj]

Adds DELETE /api/llm/endpoints/{name}, so an endpoint added by mistake or
pointing at a decommissioned provider can be removed without hand-editing the
config.

Why the referential cleanup is the substance

set_usage already refuses to create a binding naming an unregistered
endpoint, with the rationale in its own comment:

Refuse an endpoint that isn't registered: the resolver treats a dangling name
as "unset" and silently falls back, which reads as the setting being ignored.

A delete that only removed the map entry would reintroduce that exact state from
the other direction, so this clears every reference:

• openai.active promotes another registered endpoint when it named the
  removed one — null if it was the last.
• Per-usage bindings naming it are dropped, so the usage inherits the active
  endpoint again.

The whole binding is dropped rather than just its endpoint field: a binding
that kept its pinned model would silently apply that model to whichever
endpoint it inherits next.

Worth being concrete about the bug this fixes, since it is visible rather than
theoretical. Before this change, deleting an endpoint left usage_json
reporting:

{"id":"explain_code","endpoint":"alpha","model":"m-alpha","inherited":false}

inherited: false means "a deliberate per-usage choice" — pointing at an
endpoint that no longer exists. A settings UI can't distinguish that from a
valid pin, and Hoplon's picker re-inserts an unrecognised current choice (so a
hand-pinned model still displays), which resurrects the deleted endpoint as a
selectable option.

Tests

Four integration tests in management_server_tests.rs, using the existing
ephemeral-port harness:

• round-trips add → delete and returns the refreshed list
• promotes a survivor when the active endpoint is removed; active is null
  after the last one goes
• clears a usage bound to the removed endpoint (asserts inherited flips back
  to true and the pinned model goes with it)
• unknown name is 404, matching PUT /api/llm/usages/{id}

I checked the usages test actually catches the bug by reverting the retain and
confirming it fails with the dangling endpoint: "alpha", inherited: false
above, rather than passing either way.

cargo test is green (487 tests), cargo clippy --all-targets clean, cargo fmt applied.

Note on the OpenAPI description

Added the new route to docs/management-api.openapi.json. Worth flagging that
the rest of the LLM surface — /api/llm/endpoints, /api/llm/active,
/api/llm/usages, /api/llm/models, /api/llm/target — is still undocumented
there; only the Copilot login was. That is pre-existing and left alone rather
than expanded into here, but it is probably worth its own pass.

Closes #240
Unblocks ArtemisMucaj/hoplon#42