Add cache balance weight for cache-aware routing

[weireweire]

Summary

This stacked draft PR adds an optional cache_balance_weight knob to SMG cache-aware routing. A value of 0.0 preserves the existing strict best-prefix behavior. Positive values subtract a shadow-cache-size penalty from cache-match scoring, allowing common prefixes to spread across workers when the caller wants better cache footprint balance.

Changes include:

• Adds cache_balance_weight to cache-aware policy config, validation, CLI, and Python bindings.
• Adds per-tenant shadow cache footprint helpers used by cache-balanced selection.
• Keeps low-match fallback on minimum worker load.
• Adds unit coverage for default compatibility, argument plumbing, balance spreading, and strong-hit preservation.
• Updates SMG docs for the new option.

Relationship

This is stacked on pd-router-dp-aware and should be rebased onto upstream main after the DP-aware PD router change lands.

Validation

• git diff HEAD~1..HEAD --check

[weireweire]

Closing this fork-local PR; the change is being opened against sgl-project/sglang instead.