fix: V1 protocol support for multi-instance requests on unbatched models

[larmoreg]

Summary

Adds full KServe V1 protocol support to Triton's HTTP server, fixing multi-instance request handling for models with max_batch_size=0 (ensembles).

Problem

Vertex AI Batch Prediction sends requests in V1 instances format ({"instances": [...]}). When multiple instances are batched together and sent to an ensemble model with max_batch_size=0, the V1 handler incorrectly constructs input tensor shapes — prepending a batch dimension [N, 1] instead of concatenating along the first dimension [N].

This breaks Vertex AI Batch Prediction with batchSize > 1 for any ensemble pipeline.

Changes

• V1 predict endpoint (/v1/models/{model}:predict): Parse V1 instances format, convert to Triton inference requests, return V1 predictions format responses.

• Base64 binary encoding: Support {"b64": "..."} encoding for TYPE_BYTES inputs/outputs per TF Serving V1 protocol. Enables binary data (images, serialized tensors) through JSON.

• Multi-instance shape fix: For max_batch_size=0 models, concatenate instance data along the first dimension instead of prepending a new batch dimension.

• Vertex AI integration: Route /v1/models/* requests through the Vertex AI server handler.

• Tests: V1 protocol integration tests for the Vertex AI endpoint.

Tested

• Single-instance V1 predict with binary (base64) and numeric inputs
• Multi-instance V1 predict (batchSize=2) through Vertex AI Batch Prediction
• Ensemble models with max_batch_size=0 containing sub-models with max_batch_size >= N