feat: Add LoRA distribution across worker endpoints

[AmeenP]

Summary

Implements selective LoRA distribution to control which worker pods load each LoRA adapter. This enables horizontal scaling of LoRAs across multiple vLLM workers without exceeding per-worker memory/max-loras limits.

Problem

Current behavior loads ALL LoRAs on ALL vLLM workers. This doesn't scale when:

1. Number of LoRAs exceeds --max-loras capacity per worker
2. Memory becomes constrained with many LoRAs loaded

Solution

• Add DistributionSpec to DynamoModel CRD with strategies: all, fixed, percentage
• Implement deterministic endpoint selection using rendezvous hashing (HRW)
• Controller loads LoRA only on selected target endpoints
• Cleanup unloads only from target endpoints

Changes

• dynamo_model_types.go - Added DistributionSpec, TargetEndpoints, AvailableEndpoints, helper methods
• dynamo_model_controller.go - Integrated SelectTargetEndpoints() for selective loading
• discovery.go - Fixed endpoint deduplication bug (same pod appearing twice from multiple EndpointSlices)
• selector.go (new) - Deterministic endpoint selection using FNV-1a consistent hashing
• selector_test.go (new) - Comprehensive unit tests

Usage Example

apiVersion: nvidia.com/v1alpha1
kind: DynamoModel
metadata:
  name: my-lora
spec:
  modelName: Qwen/Qwen2.5-1.5B-Instruct:my-adapter
  baseModelName: Qwen/Qwen2.5-1.5B-Instruct
  modelType: lora
  source:
    uri: s3://bucket/path/to/lora
  distribution:
    strategy: fixed
    replicas: 2  # Load on exactly 2 workers

Backward Compatibility

• Default strategy is all - existing DynamoModels work unchanged
• Only changes behavior when distribution spec is explicitly set

Test Plan

E2E Testing (3 workers)

DynamoModel	Strategy	Config	Loaded On	Expected	Result
guanaco-lora	all	-	All 3 workers	All 3	✅
test-distribution-fixed	fixed	replicas=2	2 workers	2 of 3	✅
test-distribution-percentage	percentage	34%	1 worker	1 of 3	✅

Verified

• Distribution correctly limits LoRA loading to N of M workers
• Deterministic selection stable across reconciliations
• Inference works on distributed LoRAs
• Cleanup only unloads from target endpoints
• Deduplication fix working (no duplicate endpoints)
• Unit tests pass