Add KV-backed agent context checkpoints

[FabioMalpezzi]

This adds an explicit context tool for ds4-agent checkpoints and restores, backed by persisted KV state.

The scope is intentionally narrow: let the agent save a known-good model state, restore it later, and report what happened in a form that can be tested.

Included:

• agent_context checkpoint, restore, list, show, and drop.
• Restore guardrails: side-effect epoch checks, active bash-job checks, metadata/token mismatch handling, and expected/actual restore metrics.
• More robust checkpoint metadata parsing and explicit tool-action enums.
• A compaction canary e2e test that fails closed if the model returns an unusable summary.
• A KV benefit benchmark that measures avoided prefill tokens after restore.
• An adaptive self-improvement e2e test. If the native Git tool is available, the prompt asks the agent to use it for status/diff inspection. If not, the test falls back to the existing bash path with git status --short and git diff.

The self-improvement e2e is deliberately a controlled temporary-repository test. It proves the loop shape: inspect, edit, test, diff, checkpoint, restore, retest, and record evidence. It does not claim that DS4 found and optimized DS4 itself. The docs call out the stronger follow-up: a DS4-on-DS4 optimization loop that selects a small measurable improvement, implements it, runs the relevant benchmark/e2e check, and records whether the metric improved.

This PR is independent from #250. It does not add or link the native Git tool implementation. If #250 lands first, the adaptive e2e can exercise native Git inspection; otherwise it keeps working with the current bash-based path.

Tests run after rebasing on current origin/main:

• make test
• make test-agent-context-compact-canary
• make test-kv-cache-benefit
• make test-agent-context-self-improvement

KV benefit sample from this run:

• full prefill tokens: 2905
• restored prefill tokens: 18
• saved prefill tokens: 2887
• quality guard: logits equivalence