Refactor builder.py into cohesive private modules

[eric-heiden]

Part of #4191. No public API or functionality changes; the shared compatibility requirements in the tracker apply.

newton/_src/sim/builder.py combines construction, custom attributes, composition, graph transformations, validation, and finalization in one large class. This issue owns the refactoring of that file as a whole. The sections below are implementation steps, not separate issues.

Scope

• Keep ModelBuilder, its public methods/signatures/docstrings, nested public configuration types, supported subclass dispatch, and observable state compatible.
• Extract substantial rigid/joint, shape/mesh, deformable, custom-attribute/actuator, collision-filter, and validation implementations into cohesive private modules. Keep short convenience methods and authoritative state easy to find.
• Extract replication, world/builder composition, merge validation/caches, and custom-attribute/actuator merging. Preserve compact collision filters, shared geometry, source/destination aliasing, group ranges, reference metadata, and cache lifetime.
• Split fixed-joint collapse into graph discovery, reconstruction, domain fixups, reference remapping, and results. Preserve entity order, loops/disconnected components, world/articulation membership, sparse custom values, cable ranges, and mimic/equality/tendon references.
• Extract finalization into named phases for validation/world starts, model creation, particles/geometry, deformable topology, bodies/joints, counts, actuators, and custom attributes. Isolate mesh/SDF/heightfield/AABB/edge-resource preparation.
• Keep temporary phase data local to a call. Preserve allocation/execution order, device scope, geometry deduplication, existing builder mutations, and resource lifetime.
• Retain lazy importer entry points; avoid new import cycles, dynamic method installation, and stateful mixin hierarchies.

Validation

• Compare builder/model counts, labels, start/sentinel arrays, dtypes/layouts/values, custom namespaces, requested attributes, and returned remapping dictionaries before/after extraction.
• Exercise empty and mixed models, repeated finalize, non-empty composition destinations, zero/one/many replicas, custom schemas, and supported subclasses.
• Cover fixed-joint collapse, procedural rigid/deformable construction, collision filtering, and validation switches; preserve warnings/errors and their caller attribution.
• Run relevant builder/model, replication, custom-attribute, geometry/deformable, and importer tests on supported CPU/CUDA/requires_grad paths.
• Check representative setup time and allocation behavior, particularly replication, shared geometry, and deferred SDF generation.

Land coherent extractions in a sequence of reviewable PRs under this issue; no algorithm changes or hard LOC quota. Consolidates the scopes previously tracked in #4193 and #4194.