feat: provider capability detection to prevent unsupported response_format / tool_call failures

[tcconnally]

Problem

CrewAI's BaseLLM layer doesn't expose which features a provider supports (structured output, tool calling, reasoning, streaming). When a user configures response_format or uses tools with a provider that doesn't support them, the failure happens at the provider API level with a cryptic error:

• DeepSeek: doesn't support response_format — user gets raw API error
• Cerebras: thinking/reasoning behavior differs from OpenAI
• Local models via Ollama: capabilities vary by model and quantization

This is related to PR #6171 which hardcodes {"deepseek"} as a provider that doesn't support response_format. This per-provider hardcoding approach doesn't scale.

Proposed Solution

1. Add a ProviderCapabilities dataclass that each provider populates:

   @dataclass
   class ProviderCapabilities:
       supports_response_format: bool = True
       supports_tool_calling: bool = True
       supports_reasoning: bool = False
       supports_streaming: bool = True
       supports_image_input: bool = False

2. In BaseLLM, check capabilities before making API calls and raise clear error messages like: "Provider deepseek does not support response_format. Use result_as_string=True or switch to a provider that supports structured output."

3. Surface capabilities in the provider documentation

Impact

• Prevents runtime crashes from unsupported features
• Clear, actionable error messages instead of raw API errors
• Scalable approach vs. per-provider hardcoded checks (ref: PR fix: skip response_format for Deepseek (provider doesn't support it) #6171)
• Helps users choose appropriate providers for their use case