feat: cache parsed member indexes as msgpack to skip YAML re-parse on rebuild

Warm rebuilds of virtual repos (member caches valid, virtual TTL expired)
previously re-parsed all member index.yaml files on every rebuild. With 19
Helm members totalling 14 MB, YAML parsing was 60% of merge time (~6.3s of
~9.6s). Parsing each member's YAML also produces msgpack and stores it in S3
alongside the raw index. Subsequent rebuilds load the compact msgpack and skip
YAML parsing entirely.

Before: warm rebuild ~9.6s (CSafeLoader baseline)
After:  warm rebuild ~5.9s (38% faster, merge=4.7s down from ~9.6s)

pyproject.toml

@@ -14,6 +14,7 @@ dependencies = [
     "lxml>=4.9.0",
     "prometheus-client>=0.19.0",
     "python-multipart>=0.0.6",
+    "msgpack>=1.0.0",
 ]
 requires-python = ">=3.11"
 readme = "README.md"