fix: normalize provider tool choice handling and harden tool call dec…

[zblanco]

This PR tightens ReqLLM's tool-calling contract handling across OpenAI-compatible providers and hardens response decoding for malformed tool arguments.

The main changes are:

• Normalize tool_choice handling for OpenAI-compatible providers so ReqLLM inputs like :auto, :none, :required, and %{type: "tool", name: ...} are encoded consistently for provider APIs.
• Fix providers that were previously too narrow or inconsistent in this area:
  • Cerebras
  • Groq
  • Zenmux
  • Google Vertex OpenAI-compatible
  • Amazon Bedrock OpenAI
• Preserve OpenAI-style function-specific tool_choice objects for Cerebras instead of collapsing them back to "auto".
• Pass through parallel_tool_calls for Cerebras.
• Harden default OpenAI-format response decoding so tool calls with invalid or non-object JSON arguments do not crash or produce bad decoded shapes.
• Preserve malformed raw JSON where needed so ReqLLM's existing structured-output JSON repair path still works.

This came out of a broader provider-contract pass rather than only fixing the original crash case. In particular, the PR addresses:

• atom tool_choice values flowing into OpenAI-compatible providers
• forced tool-choice encoding for providers expecting OpenAI's function object shape
• inconsistent wrapper behavior in OpenAI-compatible adapters that build request bodies through shared defaults
• response-side handling of malformed/scalar tool arguments

I also added regression coverage for these provider and decoding paths, including live manual verification against real Groq and Cerebras APIs in local IEx.

Type of Change

• Bug fix (non-breaking change fixing an issue)

Breaking Changes

N/A.

Testing

• Tests pass (mix test)
• Quality checks pass (mix quality)

Additional validation performed:

• Added/updated provider regression tests for Groq, Cerebras, Zenmux, Google Vertex OpenAI-compatible, Bedrock OpenAI, and default OpenAI-style decoding
• Manually tested tool-calling with:
  • groq:llama-3.1-8b-instant
  • cerebras:gpt-oss-120b

Observed live behavior:

• Groq correctly handled tool_choice: :auto
• Groq correctly handled forced %{type: "tool", name: ...}
• Cerebras correctly handled tool_choice: :auto
• Cerebras correctly handled forced %{type: "function", function: %{name: ...}}
• Both providers returned decoded ReqLLM tool calls with expected arguments and finish_reason == :tool_calls

Checklist

• My code follows the project's style guidelines
• I have updated the documentation accordingly
• I have added tests that prove my fix/feature works
• All new and existing tests pass
• My commits follow conventional commit format
• I have NOT edited CHANGELOG.md (it is auto-generated by git_ops)

Related Issues

Closes #597