Keep duplicate devserver evaluators addressable

[Ritwij Aryan Parmar (RitwijParmar)]

Summary

Fixes #366 by making duplicate eval_name registrations in the dev server addressable instead of silently collapsing to the last evaluator.

What changed:

• build the dev-server evaluator registry by stable evaluator id rather than raw eval_name only
• keep existing behavior for unique names: /list keys and /eval { name } continue to work
• when duplicate names exist, /list exposes disambiguated keys like shared-project#1 / shared-project#2 plus id and original name metadata
• /eval accepts the new id field, also accepts the disambiguated /list key as name, and returns a 409 with candidates when the original name is ambiguous

This avoids the startup-time raise/warn stopgap and keeps both evaluators reachable.

Tests

• PYTHONPATH=py/src .venv/bin/python -m pytest py/src/braintrust/devserver/test_server_integration.py::test_devserver_keeps_duplicate_eval_names_addressable py/src/braintrust/devserver/test_server_integration.py::test_eval_falls_back_to_evaluator_project_id_when_request_omits_or_empty_it py/src/braintrust/devserver/test_server_integration.py::test_eval_request_project_id_overrides_evaluator -q
• PYTHONPATH=py/src .venv/bin/python -m compileall -q py/src/braintrust/devserver/server.py py/src/braintrust/devserver/schemas.py py/src/braintrust/devserver/test_server_integration.py
• git diff --check