docs: runnable and tested Ollama budget example

[cport1]

Summary

Completes the remaining provider-example acceptance in WebDecoy/app#1464.

• Runnable non-streaming local Ollama AsyncClient.generate integration around the existing per-attempt budget API.
• Fixed local model, raw input, bounded input/context/output settings, explicit observe/fail-open configuration and no paid fallback.
• Final response normalization, retained unknown usage, cancellation/error semantics and preserved output on settlement failure.
• Pin Ollama 0.6.3 only for the example and development tests; no runtime dependency/API/version changes.
• Explain application authentication/admission placement and deferred streaming/LangChain/LangGraph coverage.

Validation

• All 64 SDK tests pass; lint and formatting pass; wheel and sdist build.
• Seven new tests exercise the actual Ollama SDK HTTP serialization/parsing against in-memory responses: reserve-before-inference, denial-before-provider, privacy, missing/mismatched usage, explicit zero, cancellation/error and settlement outage without retry.
• Fresh virtualenv with published webdecoy-ai-protection==0.1.0a1 and ollama==0.6.3 passes all seven example tests.
• No model execution, paid calls, provider billing validation or production changes. Provider token settings remain provider-enforced, not a billing guarantee.