fix(cli): distinguish GitHub lookup failure from empty submissions (#1492)

[minion1227]

Closes #1492

Problem

find_prs_for_issue collapsed the GraphQL failure sentinel (None from
_search_issue_referencing_prs_graphql on rate limit / network / GraphQL
errors / missing issue payload) into [] via return prs or []. Downstream,
gitt issues submissions --json emitted a successful response with
submission_count: 0 — indistinguishable from a genuinely empty list. Bounty
submitters and --json automation get a false negative when GitHub lookups
actually fail.

Fix

• find_prs_for_issue now returns Optional[List[PRInfo]], propagating None
  on lookup failure (including the exception path), consistent with the
  solver_lookup_failed signaling in find_solver_from_closure_event /
  check_github_issue_closed. A falsy token still returns []
  (precondition-not-met, not a transient failure).
• fetch_open_issue_pull_requests propagates the None sentinel.
• issues submissions surfaces it via handle_exception as a
  github_lookup_failed error (JSON + human modes) instead of an empty list.
• Typed handle_exception as NoReturn so callers narrow correctly.

Tests

• None/exception lookup failures now assert None (not []).
• Added CLI coverage for the JSON and human failure paths.
• ruff clean, pyright 0 errors, 320 tests pass.

[minion1227]

This closes #1492 by giving --json consumers a stable, machine-readable github_lookup_failed error type — the whole point of the issue is letting automation distinguish a failed lookup from a genuinely empty submission list, which a generic CLI error type can't do.

Beyond the shared find_prs_for_issue sentinel change, this PR also:

• Covers the human-mode path — a failed lookup errors out instead of printing the misleading No open submissions available message used for a genuinely empty list.
• Types handle_exception as NoReturn, so pyright narrows pull_requests correctly after the failure branch.
• Handles the None sentinel at the command layer. fetch_open_issue_pull_requests has a single caller, so this adds no fan-out cost.

Tests cover both the JSON (error.type == github_lookup_failed, no submission_count) and human failure paths. ruff clean, pyright 0 errors, 320 tests pass.

[anderdc]

Cleanest fix for #1492: propagates the None failure sentinel and surfaces it as a typed github_lookup_failed error in both JSON and human modes. Verified the only caller path and tests.