Migrate resource_usage get_user_kernel_statistics_batch callers off Valkey

## Background

Three user/group resource-usage queries still read live kernel statistics from Valkey via ValkeyStatClient.get_user_kernel_statistics_batch. With kernel metrics migrating to Prometheus, these callers must be cut over so we can finally drop ValkeyStatClient and its kernel-stat keys (tracked by BA-4051).

## Target call sites
- src/ai/backend/manager/models/resource_usage.py:504 — get_user_kernel_statistics_batch in fetch_resource_usage helper
- src/ai/backend/manager/repositories/group/db_source/db_source.py:409 — group resource-usage query
- src/ai/backend/manager/repositories/user/repository.py:439 — user resource-usage query

All three follow the same pattern: collect kernel_ids → fetch live stats from Valkey → merge into the result. Pick one of two strategies (both acceptable):
1. DB-only path: drop the Valkey lookup and rely on kernels.last_stat already persisted in PostgreSQL.
2. Prometheus path: query ContainerMetricService / MetricRepository for the same fields and merge.

Decide the strategy up front and apply it consistently across all three call sites.

## Scope
- Replace each get_user_kernel_statistics_batch call with the chosen alternative.
- Remove now-unused imports of ValkeyStatClient / valkey_stat_client from the three files.
- Update any helper functions in models/resource_usage.py that exist only to merge the Valkey payload.

## Success Criteria
- [ ] None of the three files reference get_user_kernel_statistics_batch
- [ ] Resource-usage GraphQL queries for user/group return cpu/memory/io fields equivalent to current behavior (verified manually against a running session)
- [ ] grep -rn "get_user_kernel_statistics_batch" src/ai/backend/manager/ returns only the ValkeyStatClient definition (no callers)
- [ ] pants check passes for manager package
- [ ] pants test passes for affected packages

JIRA Issue: BA-6114

Start by reading the three named call sites in resource_usage.py, group/db_source.py, and user/repository.py, then inspect ContainerMetricService, MetricRepository, and kernels.last_stat to choose the DB-only or Prometheus strategy. Apply that choice consistently, remove obsolete helpers and imports, and verify with grep, the affected GraphQL resource-usage queries, pants check, and pants test.

Written by the indexing model from the issue text.