Add weighted round-robin strategy

[l-lesetong]

Currently, subscription selection seems to use normal round-robin or equal distribution.

However, different subscriptions GPT may have different quotas, costs, performance, or stability. It would be useful to assign different weights to subscriptions so that traffic can be distributed proportionally.

[bagaking]

I think this feature would be useful as a first-class routing option, with a few explicit semantics so it works predictably with the existing scheduler:

• Add routing.strategy: weighted-round-robin.
• Add per-credential selection-weight for Codex, Claude, Gemini, Vertex-compatible, and OpenAI-compatible credentials.
• Default weight should be 1; 0 should drain new selections without disabling the credential.
• Existing priority should still win first: choose the highest available priority bucket, then apply weights inside that bucket.
• For mixed-provider routing, the provider share should be based on the sum of selectable credential weights, not a fixed 50/50 provider rotation. This would also address the fairness problem described in 轮询策略有误 #2344.
• With session-affinity enabled, weights should affect new session bindings only. Existing healthy sticky bindings should remain pinned; failover/rebind should happen only when the sticky auth is unavailable, cooling down, disabled, quota-exhausted, or already tried in the current request. This should align with Improve session affinity failover for failed sticky auths #3613.
• Management API / TUI / plugin APIs should expose and preserve the weight field so it can be edited without hand-editing auth files.

Acceptance tests I would expect:

1. Weight ratio is respected within one provider and one priority bucket.
2. selection-weight: 0 prevents new selections but does not mark the auth disabled.
3. Higher priority buckets still override lower priority weighted buckets.
4. Mixed-provider routing uses credential weight sums rather than equal provider slots.
5. Session affinity keeps healthy existing bindings and only uses weighted selection for new or failed-over bindings.