[Feature] Allocation-sensitive 2D query APIs

[vmarcella]

Overview

Add non-allocating or caller-owned-buffer variants of 2D spatial queries so
hot gameplay paths can avoid per-query heap allocations.

Current State

query_point() and query_aabb() return Vec<RigidBody2D>, which is simple
and appropriate for most users, but not ideal for performance-sensitive loops.

Scope

Goals:

• Add query variants that write into caller-owned buffers or invoke callbacks
• Preserve the existing ergonomic Vec-returning methods
• Keep the public API body-oriented and backend-agnostic

Non-Goals:

• Removing the current convenience APIs
• Query-side filtering if handled in a separate issue
• Unsafe APIs in the first iteration

Proposed API

impl PhysicsWorld2D {
  pub fn query_point_into(
    &self,
    point: [f32; 2],
    out: &mut Vec<RigidBody2D>,
  );

  pub fn query_aabb_into(
    &self,
    min: [f32; 2],
    max: [f32; 2],
    out: &mut Vec<RigidBody2D>,
  );

  pub fn query_point_each(
    &self,
    point: [f32; 2],
    visit: impl FnMut(RigidBody2D),
  );
}

Acceptance Criteria

• Non-allocating query variants are added and documented
• Existing query methods keep their current semantics
• Compound-body deduplication is preserved in all variants
• Benchmarks or profiling notes justify the added API surface
• Tests cover equivalent results across allocating and non-allocating
  paths

Affected Crates

lambda-rs, lambda-rs-platform

Notes

• A callback-based API may avoid exposing buffer management concerns.
• This work should be driven by profiling.