Replace Valkey-based Live Stats with Prometheus-backed Implementation

## Background

The current live statistics system uses a custom implementation where:

- Agents collect metrics via `StatContext` and store them in Valkey (Redis) with short TTLs (8-120 seconds)
- Manager serves these stats through GraphQL live_stat fields (kernel.live_stat, agent.live_stat)
- Data is serialized with msgpack and queried via `ValkeyStatClient`

Meanwhile, a separate Prometheus integration already exists(ContainerMetricService) that queries metrics from Prometheus but is not used for the `live_stat` API.

## Problem

- Duplicate data paths: Metrics are sent to both Valkey (for live stats) and Prometheus (for monitoring/dashboards)
- No historical data: Valkey-based stats are ephemeral with short TTLs, making trend analysis impossible
- Maintenance burden: Two separate metric pipelines to maintain
- Inconsistency risk: Valkey stats and Prometheus metrics may diverge

## Proposed Solution

Replace the Valkey-based live stats system with a Prometheus-backed implementation:

1. Agent side: Remove Valkey stat publishing; ensure metrics are exported to Prometheus
1. Manager side: Modify live_stat GraphQL resolvers to query Prometheus via `ContainerMetricService` instead of `ValkeyStatClient`
1. Deprecate: Phase out `ValkeyStatClient` for statistics (keep ValkeyLiveClient for service discovery)

## Benefits

- Single source of truth: All metrics flow through Prometheus
- Historical queries: Access to time-series data for trends and analysis
- Ecosystem integration: Native compatibility with Grafana, alerting, etc.
- Reduced complexity: Eliminate Valkey stat storage and TTL management
- Lower Valkey load: Remove high-frequency stat read/write operations

JIRA Issue: BA-4039