feat(query): add temporal_relation and evidence_request plan steps

[Sidra-009]

Related issue

Related to #84

Summary

Extends the local-model query-planning path with support for temporal
relations, occurrence modes, and evidence-delivery requests, per the
examples listed in #84.

• Adds OccurrenceMode (first / best / all) to SearchMomentsPlanStep.
• Adds TemporalRelationPlanStep for before/after/during queries between
  two searches.
• Adds EvidenceRequestPlanStep for requesting a board or exact
  clips/frames as evidence.
• Adds matching opt-in flags (temporal_relations_allowed,
  evidence_requests_allowed) to QueryPlanningRequest, defaulting to
  False.
• Updates _default_plan, _valid_plan, and GroundedQueryService.plan
  in query_service.py to accept and validate the new step types, with
  the same default-False opt-in behavior.

This is an internal, application-layer change: no CLI, HTTP, MCP, or
frontend surface calls the new parameters yet, and existing callers are
unaffected since all new parameters default to False.

Out of scope for this PR (left as follow-up per #84, to keep this PR
focused on one outcome):

• Recording which local model and planning-rules version produced a
  plan, for reproducibility.
• The evidence-review guardrail for generated names/biographies/
  descriptions.

Validation

uv run --no-sync ruff check .

All checks passed!

uv run --no-sync pytest tests/test_query_service.py tests/test_application.py tests/test_control_plane.py tests/test_job_contracts.py --tb=short

65 passed in 3.43s

The full pytest -q suite could not be run end-to-end locally: several
unrelated test modules fail to collect because of pre-existing missing
optional dependencies (torch, streamlit, fastapi, httpx,
cryptography, alembic, numpy, etc.) unrelated to this change, and
uv sync --extra local-worker timed out downloading torch due to a
local network issue. These gaps were present before this change as well.