[maliberty]

Adds synthetic designs whose good placement is known in advance, and tests that check gpl recovers it. Part of moving gpl's golden tests to semantic C++ tests.

gpl_design_builder is a test-only library. It builds small designs directly in odb on a fixed 20 × 20 µm core, using one master with a pin on each side, so no LEF, DEF or liberty is needed. It has four generators:

• a chain of cells between two IOs;
• a lattice of cells tied to IOs on the die edges;
• cells on no nets;
• two unrelated lattices, each anchored by an IO on an opposite die edge.

known_answer_test checks the structure of the placement. The connected designs run both with initial placement and with Nesterov placement alone.

Test	Check
ChainLinesUpInOrder	Chain order is preserved along x (Spearman > 0.95), and the chain stays at mid height.
LatticeKeepsItsShape	Columns are ordered along x and rows along y (Spearman > 0.95).
UnrelatedClustersStayApart	Each cluster stays on its anchor's side of the core, and their bounding boxes overlap by less than 10%.
DensitySpreadsUnconnectedCells	A clump in the core corner spreads evenly under uniform density: every tile of a 5 × 5 grid holds between 0.5× and 1.5× the mean.

metamorphic_test places gcd and compares the result with three changed versions of the design, allowing a 0.2% HPWL difference:

• the IO pins mirrored left to right;
• the instance order shuffled;
• the whole design translated by (1 mm, 2 mm). Cells must also land near their shifted positions.

The measured differences are 0.012–0.034%. The tolerance is tight enough to catch a density force biased to one side, which moves the mirrored HPWL by 0.7%.

I checked the tests by breaking gpl on purpose:

Change to gpl	Tests that fail
x wirelength force zeroed	the Nesterov-only connected cases and translation
x density force zeroed	the unconnected-cells test and all three metamorphic tests
density force 1.5× to the right	the unconnected-cells test and mirroring

DensitySpreadsUnconnectedCells skips initial placement. When no cell is on a net, initial placement collapses them all into one stack at the core center. There the density force is zero, and Nesterov placement fails with GPL-0305. That is tracked as a separate fix.

The PR also includes two small fixes:

• PartialOverlapOverflowsOnlyFullBins now places one cell. The test uses it so that only the fully covered bin overflows, and it asserts the hand-computed overflow, 7.5/64. With two cells, every bin the cells touched overflowed.
• place_ios_test now depends on //src/utl for the utl/Logger.h it includes, which the layering check requires.

[gemini-code-assist]

Code Review

This pull request introduces synthetic and metamorphic global placement tests for the gpl module, including a DesignBuilder helper to construct small designs directly in odb. It adds tests verifying that the placer recovers expected structures (chains, lattices, and clusters) and spreads unconnected cells, as well as metamorphic tests comparing placement quality under mirroring, translation, and instance shuffling. Feedback is provided regarding a potential iterator invalidation issue in metamorphic_test.cpp when destroying boxes while iterating over pin->getBoxes().