[XudongZhang-1224]

Description

Fix two silent-degradation bugs in the GenAI observability mapping layer
(dashscope/finetune/reinforcement/component/observability/genai/), both
observed in production RFT agentic-rollout traces and still present on
main @ v2.0.1:

1. _bind_tool_arguments calls typing.Dict(bound.arguments); typing.Dict
   cannot be instantiated, so every @observe_tool call raises TypeError,
   which the broad except Exception swallows and replaces with
   {"_error": "could_not_bind_arguments"} as the span's tool-call arguments.
   sig.bind_partial itself succeeds; only the dict conversion crashes.
   Fix: use builtin dict().
2. openai_completion_to_output_messages and
   dashscope_response_to_output_messages only map content and
   tool_calls; reasoning_content is never read, although the schema
   (opentelemetry.util.genai.types.Reasoning, semconv ReasoningPart)
   supports it and the trace UI renders it. Fix: export Reasoning from the
   _core facade and insert a Reasoning part first in each output message.

Related Issue: N/A — no upstream issue; reported from production traces
(dashscope 1.26.7, and live-confirmed on 1.27.6 in a running FC function).

Security Considerations: This change adds no logging of secrets and
exposes no credentials. (An unrelated pre-existing issue is noted under
Additional Notes.)

Type of Change

• Bug fix
• New feature
• Breaking change
• Documentation
• Refactoring

Component(s) Affected

• Model
• Application
• Common
• Documentation
• Tests
• CI/CD

Checklist

• Pre-commit hooks pass
• Tests pass locally
• Documentation updated (if needed)
• Ready for review

Testing

• Unit: against this source tree, _bind_tool_arguments returns the real
  kwargs (no sentinel); the OpenAI converter emits parts
  [reasoning, tool_call] for a completion carrying reasoning_content;
  the DashScope converter emits [reasoning]; a message without reasoning
  stays [text] (no regression).
• Integration: with an in-memory SpanExporter and tracing enabled, the
  decorated tool path records gen_ai.tool.call.arguments as the real
  command JSON (no sentinel) plus gen_ai.tool.call.result; the decorated
  LLM path (demo-style result_mapper) records gen_ai.output.messages
  containing reasoning and tool_call parts.
• Platform: an agentic-rl job running a wheel built from this branch
  (FC startup log: Successfully installed dashscope-2.0.1) shows real
  shell commands in bash tool span inputs and reasoning parts in chat span
  outputs in the trace UI.
• The repository pytest suite has no coverage for this module; maintainer-side
  CI expected to run as usual.

Additional Notes

• Compatibility: when loongsuite-util-genai is absent, _core.Reasoning
  is None and the new helper no-ops, preserving the module's
  silent-degradation contract; messages without reasoning_content are
  byte-identical to before.
• Separate pre-existing issue observed while validating (not addressed here):
  the AgenticRL submit-path log prints the full job request payload at INFO,
  including component env vars that carry API keys. Suggest masking env
  values in that log in a follow-up.