[DYNAMO] add chat nvext transport

[AmeenP]

Summary

Adds the Dynamo token-in wire shape to the standalone renderers package, which is now the canonical home for renderer client code.

• adds RendererTransport = Literal["vllm", "dynamo"]
• keeps the existing vLLM /inference/v1/generate behavior as the default transport="vllm", including current multimodal features support from main
• adds transport="dynamo", posting placeholder chat messages plus nvext.token_data to /chat/completions
• sends renderer stop token IDs through the standard chat stop field for Dynamo
• requests nvext.extra_fields: ["completion_token_ids"] and reads generated token IDs from Dynamo's token_ids / nvext.completion_token_ids response locations
• exports the transport type from renderers.__init__

Context

This replaces the old verifiers PR path that edited packages/renderers/...; latest verifiers/main now depends on the standalone PrimeIntellect-ai/renderers repo.

Paired stack:

• ai-dynamo/dynamo#9509 minimal Dynamo nvext contract PR split from ai-dynamo/dynamo#9382
• PrimeIntellect-ai/verifiers#1287
• PrimeIntellect-ai/verifiers#1313
• PrimeIntellect-ai/prime-rl DYNAMO stack

Validation

• ruff check renderers/client.py renderers/__init__.py tests/test_client.py
• ruff format --check renderers/client.py renderers/__init__.py tests/test_client.py
• python -m py_compile renderers/client.py renderers/__init__.py tests/test_client.py
• direct execution of the pure client tests, including test_generate_can_use_dynamo_transport

uv run pytest tests/test_client.py still exits 139 on the local macOS/Anaconda environment during pytest startup capture initialization; relying on Linux CI for the full pytest run.