feat: report judge context overflow as could-not-evaluate, not score-0

[pdebjyot]

What & why

Closes the first half of #342. A judge-based evaluator whose prompt overflows
the model's context window currently records score: 0 / test_pass: False —
indistinguishable from a genuine zero-quality result. That silently drags
down overall_score and the pass rate, so a run where the judge never actually
read the trace looks like a run where the agent failed.

This makes context overflow a distinct could-not-evaluate outcome that is
excluded from aggregates instead of being counted as a quality failure.

What's in the PR

• types/evaluation.py — EvaluationStatus enum {evaluated, could_not_evaluate}.
• experiment.py — a shared _error_result helper tags only
  context-overflow errors (via the existing _is_context_exceeded predicate)
  as could-not-evaluate. Every other exception stays a real
  score: 0 / test_pass: False failure, so a genuine evaluator bug (e.g. a
  KeyError in a custom evaluator) still surfaces rather than being hidden.
  Statuses are threaded through aggregation; could-not-evaluate cases are
  excluded from overall_score and from the diagnosis on-failure trigger.
• types/evaluation_report.py — new statuses list + mean_score()
  classmethod (excludes could-not-evaluate); flatten() threads it; _display
  renders N/A for an unscored case instead of 0.00. Reports written before
  this field existed deserialize unchanged.
• display_console.py — could-not-evaluate cases are dropped from both the
  numerator and denominator of the printed Pass Rate.

Behavior

case	before	after
judge overflows context	score: 0, counted as a fail	could_not_evaluate, excluded from score + pass rate, shown N/A
genuine evaluator error	score: 0, counted as a fail	unchanged — still a fail
everything fits	unchanged	unchanged

Raw per-case scores / test_passes lists are left intact; only the aggregates
and the display change.

Tests / checks

• Report-level: mean_score exclusion, flatten threading, round-trip
  serialization, legacy reports without the field.
• Display-level: pass-rate excludes could-not-evaluate; all-unscored is guarded
  against divide-by-zero.
• Experiment-level: overflow → could-not-evaluate & excluded from score; a
  non-overflow error stays a failure.
• Full suite 1700 passed; ruff check and mypy -p src clean.