accelerator_quantum_size is a global etcd setting, not per-resource-group

## Problem

`accelerator_quantum_size` is exposed as a field on the `ScalingGroup` GraphQL type, but its resolver reads from a single global etcd key (`config/plugins/accelerator/cuda/quantum_size`) regardless of which resource group is queried. All resource groups return the same value.

Additionally, `CreateScalingGroupInput` and `ModifyScalingGroupInput` do not include this field, so it cannot be set via GraphQL mutations.

**Relevant code:** `src/ai/backend/manager/api/gql_legacy/scaling_group.py:439-441`

## Proposed Improvement

- Add a per-resource-group `accelerator_quantum_size` column to the DB (or per-group etcd key).
- Add the field to `CreateScalingGroupInput` and `ModifyScalingGroupInput` so it can be set via mutations.
- Fall back to the global etcd value when the per-group value is not set.

—

Captured while working on branch: main

JIRA Issue: BA-5692