kvbm-config: workspace test run breaks figment f64 extraction via serde_json arbitrary_precision unification

[xiaguan]

`cargo test --release --workspace --lib` (the documented unit-test command) fails on main:

test tests::test_profile_with_defaults_and_overlay ... FAILED
panicked at kvbm/kvbm-config/src/lib.rs:531:89:
InvalidType(Map, "f64") at path ["cache", "host", "cache_size_gb"]

Repro

cargo test --release -p kvbm-config --lib                       # green (63 passed)
cargo test --release -p kvbm-config -p pegainfer-server --lib   # red — same code

Cause

vllm-chat (git dep of pegainfer-server via vllm-server) enables serde_json/arbitrary_precision:

cargo tree -i serde_json -e features
├── serde_json feature "arbitrary_precision"
│   ├── vllm-chat v0.1.0 (github.com/vllm-project/vllm.git rev 4aaed4c)

Workspace feature unification turns it on for every serde_json user. Under arbitrary_precision, JSON numbers deserialize as an internal {"$serde_json::private::Number": "2.0"} map, so figment's Json::string(...) source hands kvbm-config a Map where cache_size_gb: Option<f64> expects a float (kvbm/kvbm-config/src/lib.rs:283 extract, :531 unwrap in test).

Verified on main 927c7c1; unrelated to any feature branch (single-package runs are green everywhere).

Acceptance

• cargo test --release --workspace --lib green on main.
• Either drop arbitrary_precision from the vendored vllm-chat (preferred — nothing in this workspace needs it), or make kvbm-config's figment path tolerant of Number-as-map.

[mvanhorn]

Implemented a fix in #317 - made cache_size_gb tolerant of the serde_json/arbitrary_precision number-map form via a deserialize_with helper, so the field extracts correctly under workspace feature unification without pinning/stripping the upstream feature.