AlphaEvolve implementation for RSIKit using Slick and Pydantic

This is an independent implementation of the published AlphaEvolve algorithm,
not Google DeepMind's internal agent. No AlphaEvolve agent source is published
in the official accompanying repository.

The original island founding and weaker-half reseeding in original/agent.py adapt
the algorithm in the official FunSearch implementation:
https://github.com/google-deepmind/funsearch/blob/main/implementation/programs_database.py
Copyright 2023 DeepMind Technologies Limited
Licensed under the Apache License, Version 2.0 (see LICENSE.funsearch).

Modifications: ordinary Python state and RNG; arbitrary complete source text;
per-cell, per-objective champions; caller-selected target metric; reset by
completed attempts; asyncio workers; injected evaluators; external Slick prompts.

AlphaEvolve's official results and verification code are available at:
https://github.com/google-deepmind/alphaevolve_results
They contain problem-specific results, not reusable agent code, and are not vendored.

RSIKit adaptation: Python Solution classes, Pydantic JSON proposals instead of
paper-format text diffs, syntax/interface checks, and isolated Gymnasium evaluation.
Adapted from the local slick-bits implementation; no runtime dependency on slick-bits.
The original variant preserves local baseline behavior from commit 5ba5685.
The improved variant preserves changes from commit 8a6bc96: independent island
founders instead of copying the best initial program to every island, plus
per-seed rewards as prompt feedback. Both are local adaptations; neither is
Google DeepMind's internal AlphaEvolve agent.

The paper variant independently implements the mechanisms described in sections
2.1-2.6 of https://arxiv.org/abs/2506.13131 . Its persistent per-metric/per-niche
archive, rank sampling, descriptor discretization and member migration are local
algorithm choices because those details are not specified by the paper. It reuses
the local generation and repair machinery, with structured JSON diffs rather than
the paper's textual SEARCH/REPLACE envelope. See docs/ALPHAEVOLVE.md for coverage
and experimental differences; no reproduced paper or BipedalWalker score is claimed.
