feat(eval): warn when a trajectory_id appears more than once in a run

[isarasua]

What changed and why

One trajectory_id is one conversation, but nothing checked that a run holds it once. A writer that
adds a restarted conversation's row beside the earlier one gives it two rows, and usersim eval
scores both, so it counts twice. Now eval warns, naming the ids:

2 trajectory_id(s) appear more than once in this run: t000000000000001 (2 rows), ... Each copy is
evaluated as a row of its own, so such a conversation can count more than once; a restarted
conversation's row should replace the earlier one.

• Over the whole run. The check reads every row's id before any sampling, so a pass with
  --num-records or --num-records-per-locale still sees a repeat outside its sample. It names up to
  five ids and counts the rest.
• Rows are evaluated as they are. Only the warning is new.
• Docs. docs/outputs.md states the convention: one id per conversation per run, derived from
  what the conversation stands for, not from the attempt. A restart replaces its row. It also says what eval does with a repeat.

Closes #40

How it was verified

• The new tests in tests/test_cli_eval_sampling.py:
  • a repeated id is named, with its row count, for a full pass, --num-records 1 and
    --num-records-per-locale 1, and every row is still evaluated
  • a run without repeats gives no warning
• Deliberate breaks: dropping the check from either branch of _choose_rows fails its cases.
• The full suite passes on Python 3.12 and 3.10 (3,987 tests). make check-all, make check-license-headers and
  make check-doc-links pass.

🤖 Generated with Claude Code

[3mei]

@isarasua one doc change is needed

[3mei]

Merging