Fix tool-calling resilience (retry + circuit breaker)

[imnishantmehra]

Description

Tool-calling flows were bypassing BaseProvider.complete() by calling complete_with_tools() directly in multiple internal paths, which meant tool calls didn’t benefit from the standard retry + circuit breaker + retry-metrics wrapper. This PR routes tool execution through the same resilience layer as normal completions, while remaining compatible with existing provider stubs/mocks used in tests.

Type of Change

• Bug fix
• New feature
• Documentation update
• Performance improvement

What Changed

• BaseProvider:
  • Auto-detects tool support when a provider implements complete_with_tools(...).
  • Implements _complete_with_tools_impl(...) to dispatch to complete_with_tools(...) (normalizes messages/system prompt) so tool calls run through _execute_with_retry(...).
  • Passes tool_choice through as-is (supports object/dict-style tool forcing).
• Internal call sites:
  • Updated tool-call paths in cascadeflow/agent.py, cascadeflow/core/cascade.py, and cascadeflow/streaming/tools.py to go through the retry/CB path when the provider is a real BaseProvider.
  • Keeps compatibility with non-BaseProvider provider stubs by falling back to complete_with_tools(...) where appropriate (test/mocks).
• Provider config hardening:
  • VLLMProvider normalizes host-only base URLs to include /v1 (prevents misconfig and fixes failing tests).
• Tests:
  • Added deterministic unit test proving tool calls are retried: tests/test_provider_tool_retry.py.
  • Made a couple provider env tests deterministic via clear=True where needed.

Testing

• Added new tests
• pytest tests/test_provider_tool_retry.py -v
• pytest tests/ -v -m "not integration and not requires_api and not requires_ollama and not requires_vllm"
• Ran formatting script