[ty] Cache non-terminal-call reachability prefixes (#26810)
## Summary
Statement-level call predicates are assigned in source order but appear
in reverse order in reachability decision diagrams. We warm preceding
calls in source order to avoid a deeply recursive Salsa query chain, but
the existing implementation rescans the entire predicate prefix for
every later expression-inference query. With `n` statement calls, the
query for call `i` scans the first `i` predicates, producing `1 + 2 +
... + n = O(n²)` total scanning even though the individual call analyses
are cached.
This change indexes statement-call predicates once for scopes with more
than 16 calls and decomposes each requested prefix into reusable
power-of-two ranges. Each leaf analyzes 16 actual calls iteratively,
while smaller scopes retain the direct path without allocating the
tracked index.
This revisits the approach from #26763, using call-based rather than
raw-predicate-based chunks so the cached work remains proportional to
the expensive predicates.
This is the first of three related PRs.
[#26811](https://github.com/astral-sh/ruff/pull/26811) caches sparse
reachability checkpoints, and
[#26793](https://github.com/astral-sh/ruff/pull/26793) defers
statement-call narrowing gates.
