github: Batch graphql requests to avoid resource limit errors

[f0rmiga]

Problem

With a large enough topic stack, revup upload fails outright with GitHub graphql errors — intermittently Resource limits for this query exceeded or Timeout on validation of query.

Two call sites build a single request that grows linearly (or worse) with stack size:

• Github.query_everything() aliases one pullRequests(headRefName:) lookup per head ref into one query. Since every topic in a relative chain also adds its relative branch to the lookup list, a 41-topic stack produces ~80+ aliases, each nesting reviews, assignees, labels, timeline items, and comments.
• create_pull_requests() / update_pull_requests() alias one mutation per PR into one request. These hit the limit even sooner: each PR's body and review-graph comment lists the whole relative chain, so the payload grows quadratically with stack size.

Fix

Batch the requests instead of aliasing everything into one:

• query_everything() looks up head refs 10 per query (MAX_HEAD_REFS_PER_QUERY); users/labels/teams ride along with the first batch. Results are merged in order, so callers are unaffected.
• create_pull_requests() / update_pull_requests() send mutations 5 PRs per request (MAX_PRS_PER_MUTATION), smaller because of the body/comment payloads.

The public methods keep their signatures (the original bodies became _impl methods), and stacks below the batch sizes send byte-identical requests to before, so there is no behavior change for small stacks. Batches are sequential; they could be issued concurrently with asyncio.gather if you prefer — happy to change it.

Verification

• Reproduced on a 41-topic relative-chain stack in a large monorepo: the single-query form fails reproducibly during real uploads, while the batched form completes and updated all 41 PRs.
• Stress-tested query_everything() read-only against 2,173 head refs (218 batches) without errors, and smoke-tested this branch's code against the real API.
• make lint (mypy, ruff, pylint) and make test (201 passed) are clean.