fix(backup): prevent duplicate predicates in backup manifest for vector indexes

[shiva-istari]

Description

Fixed a bug in ProcessBackupRequest where vecPredMap[gid] was being built using predMap[gid] as the append base instead of vecPredMap[gid]. This caused all base predicates in the group to be duplicated in the manifest every time a vector predicate was found, making manifest.json grow larger and larger with each backup.
With multiple vector predicates in the same group, each iteration also overwrote the previous result, so only the last vector predicate's supporting entries (__vector_entry, _vector, __vector_dead) were kept the rest were silently lost.

[matthewmcneely]

Looks good but there's no test covering the scenario where a group has multiple vector predicates with indexes. A test should:

1. Create a schema with 2+ float32vector predicates with HNSW indexes in the same group
2. Run ProcessBackupRequest
3. Verify the backup request's Predicates list includes the supporting predicates (VecEntry, VecKeyword, VecDead) for all vector predicates, not just the last one

Without such a test, this exact bug pattern (appending to the wrong map key) could easily regress.