fix(llmmultiroundrouter): increase decomposition token budget

[panpan0000]

Description

This PR fixes a routing fallback problem in llmmultiroundrouter.
For some reasoning-heavy prompts, 512 tokens is not enough for the decomposition + routing step. When the output is cut early, the router may fall back to the first model in llm_data.

Related Issues

• Fixes #
• Related to #

Type of Change

• Bug fix (non-breaking change that fixes an issue)
• New feature (non-breaking change that adds functionality)
• New router implementation
• Breaking change (fix or feature that would cause existing functionality to change)
• Documentation update
• Performance improvement
• Code refactoring
• Test addition or update

Changes Made

• add decomposition_max_tokens for llmmultiroundrouter
• change the default decomposition token budget from 512 to 2048
• update the example config and router README
• add tests for the default value and config override

Router Affected (if applicable)

• knnrouter
• svmrouter
• mlprouter
• mfrouter
• elorouter
• routerdc
• automix
• hybrid_llm
• graphrouter
• causallm_router
• router_r1
• gmtrouter
• knnmultiroundrouter
• llmmultiroundrouter
• Other / Core framework

Checklist

• I have tested my changes locally
• I have reviewed my own code
• I have added/updated documentation where necessary
• I have added tests for my changes (if applicable)
• My changes do not introduce new warnings or errors
• Any dependent changes have been merged and published

Test Plan

python3 -m pytest tests/test_llmmultiroundrouter_decomposition.py -q

Screenshots (if applicable)

Additional Notes

This is a small fix to make routing output more stable for longer decomposition prompts.