commit-reach: terminate merge-base walk when one paint side is exhausted

[spkrka]

paint_down_to_common() computes merge bases by walking the commit
graph backwards from two sets of tips, painting commits with PARENT1
and PARENT2 flags. The walk currently terminates when no non-stale
commits remain in the queue.

This termination condition forces the walk to visit commits that
cannot contribute new merge-base candidates. A new merge base can
only appear when a path carrying only PARENT1 paint meets a path
carrying only PARENT2 paint. Once one side has no exclusive commits
remaining in the queue, no new merge base can form, and the walk
can stop.

This is particularly visible in repositories that use merge
commits. Each merge in the first-parent history introduces a side
branch whose ancestry is painted with only one side. The existing
termination condition keeps the walk alive until that entire
ancestry is drained, even though the merge base was found long ago.

The same effect is triggered by repository imports. In a large
monorepo, importing a project brings in a separate history with its
own root at generation zero. Any merge-base query that straddles
an import boundary must then drain all the way to that root.

This was discussed in an earlier RFC thread [1], where Derrick
Stolee confirmed the correctness of the per-side approach and
provided the "release branch" shape as another motivating example.
Elijah Newren independently discovered the same optimization and
shared his implementation in PR #2150.

How it works

This series replaces the single max_nonstale pointer with a
paint_queue struct containing per-side integer counters (p1_count,
p2_count, pending_merge_bases). A centralized
paint_count_transition() function maintains the counters on every
flag change.

The walk terminates early when all three conditions hold:

(a) the walk has entered the finite-generation region
(generation < GENERATION_NUMBER_INFINITY),
(b) no pending merge-base candidates remain in the queue, and
(c) one side's exclusive count has reached zero.

Condition (a) is essential: in the finite-generation region,
generation ordering guarantees topological traversal, so paint on
already-visited commits is final. In the INFINITY region (commits
not in the commit-graph), commit-date ordering can violate this, so
the optimization is disabled and the walk falls back to the existing
termination condition.

Patch organization

1/6 commit-reach: introduce struct paint_queue with per-side counters
2/6 commit-reach: terminate merge-base walk when one paint side is exhausted
3/6 t6099: test merge-base with ancestor among candidates
4/6 t6600: add test cases for side-exhaustion edge cases
5/6 p6012: add perf test for merge-base with deep side branch
6/6 Documentation/technical: add paint-down-to-common.adoc

Patch 4 includes test cases from Elijah Newren's branch, with
his authorship and Signed-off-by preserved.

Benchmarks

Measured on a 2.6M-commit monorepo with commit-graph. Baseline
is v2.55-rc1.

Import boundary scenario: merge-base across a repository import
where the imported history has its own root at generation zero.

merge-base --all ABOVE BELOW_IMPORT     4.293s ->    8ms  (537x)
merge-base --all HEAD  BELOW_IMPORT     5.590s ->   85ms   (66x)
merge-tree       ABOVE BELOW_IMPORT     5.345s ->   13ms  (411x)

Merge-heavy history: merge-base --all between commits on the
same first-parent line, where merge commits create many one-sided
subtrees.

merge-base --all HEAD HEAD~1000         5.404s ->    7ms  (772x)
merge-base --all HEAD HEAD~5000         5.310s ->   14ms  (379x)

No regression for single merge-base and ancestry checks:

merge-base       HEAD HEAD~100           7ms ->    9ms   (~1x)
merge-base       HEAD HEAD~1000          7ms ->   11ms   (~1x)
merge-base --is-ancestor ~1000 HEAD      9ms ->    8ms   (~1x)

[1] https://lore.kernel.org/git/CAL71e4Ps-2_0+uuZu43N9pFnXBemoAohPs_eyRJf8taXHJPAXQ@mail.gmail.com/T/#u