Make LMP improving-aware

[KnightStackSys]

Adds an improving/worsening signal to the per-ply search stack and uses it only to tune Late Move Pruning.

Base

This branch is intentionally based on late-move-pruning, not main, so it can be tested directly against the existing LMP implementation.

Search stack

Adds SearchStackEntry, which preserves the existing HistoryContext interface and stores:

• staticEval
• improvementRate
• improving

4PC improving calculation

IronPhoenix compares static evaluation with ply - 4, not ply - 2, so the comparison is against the same player's perspective after one complete Red/Blue/Yellow/Green turn cycle.

• improving margin: +25 cp
• strongly worsening margin: -100 cp

LMP tuning

Base LMP threshold remains:
3 + depth * depth

Then:

• improving: search 2 + depth additional moves before LMP
• strongly worsening: prune 1 + depth / 2 moves earlier
• clamp threshold to at least 3

No LMP depth cap is reintroduced. Existing check, PV, TT-move, mate-window, checking-move, capture/promotion, and history safeguards remain intact.

Scope

This experiment does not modify LMR reductions yet. The improving signal affects LMP only, allowing a clean strength test.

Recommended test:
late-move-pruning vs late-move-pruning-improving with identical HCE, openings, TC, hash, threads, and MultiPV=1.

[KnightStackSys]

LMP was successful on first run