refactor: adopt SymbolicRegression abstract crossover API

[MilesCranmerBot]

Summary

This updates the LaSR v2 plugin port from the superseded SymbolicRegression.jl crossover proposal hook to the merged first-class crossover API.

• add LLMCrossover <: AbstractCrossover and implement SymbolicRegression.crossover with CrossoverResult;
• let SymbolicRegression own constraint retries, evaluation, accounting, and replacement;
• avoid repeating an expensive LLM request on constraint retries;
• configure native mutations and crossovers through Options.mutations and Options.crossovers;
• preserve LaSROptions compatibility by mapping the legacy conditional LLM crossover probability to exact LLM/subtree weights;
• pin the development source to SymbolicRegression.jl 74cf6b79, which includes the abstract crossover API and the stdin monitoring fix;
• update docs, examples, and deterministic mocked integration coverage.

This supersedes the crossover integration in #85, which depended on the now-removed propose_crossover direction.

Validation

• Julia 1.11.9, SymbolicRegression.jl 74cf6b79
• focused mocked plugin integration: 15/15 assertions passed
  • mutation/search: 8
  • concept lifecycle: 2
  • end-to-end abstract crossover search: 2
  • legacy crossover probability mapping: 3
• Aqua passed: unbound parameters, exports, project consistency, stale dependencies, compat bounds, piracy, persistent tasks, and ambiguities
• all configured pre-commit hooks passed, including JuliaFormatter 1.0.56 and Prettier
• git diff --check passed

The tests inject a deterministic local LLM function, so no credentials or external provider calls are required.

Remaining scope

• no authenticated/provider-backed LLM test was run;
• multiprocessing plugin-state serialization remains unexercised in this focused update.