resource control: align token bucket semantics under Service Limit

## Enhancement Task

Follow up on #11240 and #11244 to consistently apply the Fill Rate and Burst Limit semantics introduced by Service Limit across PD and its client. Preserve existing resource-group behavior and keep changes focused on the affected state transitions.

- [ ] Synchronize Service Limit updates and override publication to prevent stale constraints after disabling the limit; cover resource-group mode changes (#11111).
- [ ] Handle client token balances correctly when switching from unlimited to limited mode.
- [ ] Use consistent participants for demand calculation and allocation, preventing negative rate overrides and inflated burst capacity.
- [ ] Preserve refunds for tokens rejected by the service limiter, including zero-capacity limited buckets.
- [ ] Separate configured and effective Fill Rate during initialization so a zero override does not rewrite user settings.
- [ ] Restore override metric series correctly when constraints become active again.
- [ ] Add focused regressions and realistic validation for TiDB node churn, group creation/modification, priority contention, fractional allocations, and Service Limit changes, alongside normal workloads.

Related umbrella: #9296.

Start by tracing the Service Limit Fill Rate and Burst Limit state transitions across PD and its client, then read the related issues #11240 and #11244. Validate the listed transitions with focused regressions covering resource-group changes, token balances, refunds, allocations, metrics, and TiDB node churn; done means constraints and overrides remain consistent through limit changes and normal workloads.

Written by the indexing model from the issue text.