feat: add TRT-LLM, Dynamo KVBM integrations + dynamo-semblend Rust crate

Major additions for SemBlend v0.2.0:

TRT-LLM integration (semblend/integration/trtllm/):
- TRTLLMPyTorchBackend implementing SemBlendBackend ABC
- KV cache layout adapter (stride computation for TRT-LLM's paged cache)
- Model engine hook with 3 approaches (token sub, radix patch, block inject)
- SemanticCacheLookupProvider + PostPrefixLoadHook upstream ABCs
- SemBlendProvider reference implementation
- Turnkey launcher (semblend-trtllm CLI)
- 54 tests passing

Dynamo KVBM integration (semblend/integration/dynamo/):
- SemBlendKvIndexerWrapper wrapping Dynamo's KvIndexer
- SemBlendEventPublisher for NATS semantic events
- 14 tests passing

dynamo-semblend Rust crate:
- SemanticKvIndexer implementing KvIndexerInterface trait
- DonorStore with SIMD cosine similarity
- EmbedClient for MiniLM sidecar
- 16 Rust tests passing

Deploy infrastructure:
- TRT-LLM: K8s manifests, Docker Compose, engine build scripts
- Dynamo: DynamoGraphDeployment configs, SemBlend proxy
- Benchmark results: TRT-LLM baseline + Dynamo baseline + SemBlend

Benchmark results (Dynamo + TRT-LLM, 736 samples):
- NarrativeQA: 19.3% baseline → 29.3% with SemBlend (+10pp)
- Full 5-dataset baseline established for TRT-LLM and Dynamo KVBM

Signed-off-by: Zach Bennett <zach@worldflowai.com>

Author: zbennett10

deploy/dynamo/model-cache-pvc.yaml

@@ -0,0 +1,12 @@
+apiVersion: v1
+kind: PersistentVolumeClaim
+metadata:
+  name: model-cache
+  namespace: semblend-trtllm
+spec:
+  accessModes:
+    - ReadWriteOnce
+  resources:
+    requests:
+      storage: 50Gi
+  storageClassName: "gp3"