[Engine & Testing]: Reduced memory consumption to fit large sims into H100s and reduce sim compile time, slim down codebase, and avoid running complex tests

[QuentinWach]

What Changed

• The main runtime change is a compiled 3D update refactor. The branch moves away from the prior sparse-shell compiled memory mode and toward direct material/primitive-profile updates. The key pieces are explicit field-shape tracking, material conductivity/inverse-permittivity inputs, primitive CPML profile support, and dense fallback when CPML profiles are not separable.
• 3D lossy material updates now derive coefficients in the fused loop from conductivity and inverse permittivity rather than carrying large precomputed dense source grids. That is primarily a memory reduction and consistency change.
• 3D ModeSource compilation now builds compact local residual slabs instead of larger full-field delta arrays. The purpose is to keep compiled source specs smaller while preserving the exact full-minus-masked incident update behavior.
• The compiled engine was made dtype-stable after PDK/mode-solver paths enable or expose x64 values. run_compiled now initializes fields to the compiled program precision, source patches and scan carries are cast back to carry dtype, CPML profiles preserve dtype, and snapshot indices avoid int32/int64 mismatches.
• Tests were expanded around compiled/material behavior: CPML primitive profile fallback, 3D metallic edge zeroing, material coefficient equivalence, compact ModeSource residual compilation, memory reporting, and compiled/step equivalence. The expensive integration/characterization tests were left in the tree but disabled from normal collection.
• Examples and old generated artifacts were aggressively pruned. Many old examples, benchmark scripts, and docs/architecture/index.html were deleted; several remaining examples were renamed to drop numeric prefixes.

Why
The branch is mainly about reducing memory pressure and stabilizing compiled 3D FDTD execution: smaller CPML representation where possible, no sparse-shell special mode, less dense precomputation, compact ModeSource residuals, and fewer dtype/order-sensitive failures. The later test and example changes are about keeping CI/runtime costs down and removing old or expensive artifacts.