kvstore: save continued checkpoints after crossed interval frontiers

[JordiPosthumus]

Continued KV saves required live_tokens % step == 0. After restoring an
unaligned checkpoint, fixed-size prefill chunks can cross every interval
without ever landing on an exact multiple, so no continued checkpoint is
written.

For example, a production run restored 9,703 tokens and advanced in
2,048-token chunks with a 16,384-token save interval. The sequence
9703 + n * 2048 never reaches an exact interval boundary.

This patch saves at the first valid live position at or beyond each crossed
frontier. It also resets the per-slot frontier to the checkpoint actually
restored, including backward restores and rebuilds from zero, and performs
interval rounding in int64_t before checking the int range.

The cache format, stored tokens, graph payload, and defaults are unchanged.

Tests cover crossed intervals, forward/backward/zero restores, and overflow
at INT_MAX. ds4_test --server, ds4_agent_test, forced Metal and CPU
builds, and the server suite under ASan+UBSan pass on Apple Silicon. No model
is required for these scheduling/state regressions.

Current-main recheck — 2026-09-03

Applied independently to main b0a147a7fba6d1a104d047d5a140e9bb4bfc13cd in a clean
worktree. The default build, ds4_test --server, and ds4_agent_test pass.
The change remains independently useful; cancellation rollback's save-counter
repair is handled separately by #960.