Fix future() negation edge cases; add prune_empty_branches (#269)

[jtdub]

Summary

Closes #269 (next side; a master port follows as a separate PR).

compute_future now resolves negations in an explicit, documented order:

1. Exact positive match → the negation removes the line and neither survives (case 1). The root cause was ordering: the idempotency check ran first and EOS's unanchored neighbor \S+ description rule matched the negation line itself, keeping it as a literal child while displacing the original.
2. Idempotency rules → unchanged replace semantics, deliberately including rule-tracked negated forms: IOS no logging console still replaces logging console X and persists in the render — real device behavior for display-default commands, and the existing remediation roundtrip depends on it. A stale-valued EOS negation likewise displaces the tracked line but stays visible, preserving the did-not-apply-cleanly signal the issue explicitly asked to retain.
3. Shorthand prefix match → no description removes description foo (and no X removes all X ... lines), as devices do (case 2).
4. Unmatched negations are kept — both for no ... lines native to running configs and as the apply-signal.

Case 3 ships as an opt-in: HConfig.future(change, prune_empty_branches=True) prunes sections the change emptied out (cascading upward, e.g. XR router static / address-family), while sections that were already empty — or newly added empty — are kept, so legitimately-empty interfaces don't vanish.

Test plan

• Six new integration tests: all three issue repros verbatim (adapted to v4 constructors), the stale-value idempotency-replace pin, the unmatched-negation signal pin, and prune on/off/originally-empty behavior.
• Full suite passes (702 tests) — including the IOS logging console roundtrip that constrains the idempotency-replace semantics.
• poetry run ./scripts/build.py lint-and-test — all linters green.

🤖 Generated with Claude Code