feat(llm): add configurable GenerationParams across agent runners

[Sagar-CK]

Description

Standardizes LLM generation configuration across all agent runners and the LiteLLM backend. Replaces hardcoded values (e.g. max_tokens=2048) with a shared, env-driven GenerationParams dataclass and per-SDK mappers.

Problem

• LiteLLMBackend hardcoded max_tokens=2048 with no override path
• No way to enable/disable extended thinking / reasoning consistently
• temperature was the only tunable param; no shared place for top_p, stop, etc.
• SDK runners (claude-agent, openai-agent, deep-agent) had no shared generation config

Solution

• Introduce GenerationParams (src/llm/generation.py) with env defaults via LLM_* vars
• Thread through LLMBackend.generate(..., *, params=) with merge semantics (env → ctor → per-call; temperature= always wins)
• Map to each SDK: LiteLLM kwargs, ClaudeAgentOptions (effort/thinking), OpenAI ModelSettings, LangChain ChatOpenAI
• Gate reasoning by model family heuristics; strip + warn when unsupported (e.g. WatsonX)
• Add unit tests for env parsing, merge semantics, and litellm kwargs matrix

Env vars (documented in .env.public):

• LLM_MAX_TOKENS (default: 4096)
• LLM_TEMPERATURE (default: 0.0)
• LLM_TOP_P, LLM_STOP, LLM_REASONING_EFFORT, LLM_THINKING_BUDGET_TOKENS

Type of Change

• New Benchmark Scenario (Industry/Asset type)
• Evaluation Metric / Scorer
• Agentic Orchestration Logic (ReAct, Plan-Execute, etc.)
• Infrastructure / Tooling Improvement

Related Issues

• Refs: Standardize LiteLLM params: thinking toggle, max_tokens, and generation settings #304

Testing & Validation

• Unit Tests: uv run pytest src/ -v -k "not integration" — 357 passed

[DhavalRepo18]

@Sagar-CK It seems that too many changes are coming along.

This is potentially due to formatting. Will it be okay if you disable formatting for this PR, and maybe we can make a pass to reformat all the files with standard formatting?

[DhavalRepo18]

Submit code without Python code formatting.

[DhavalRepo18]

@Sagar-CK I left few comment.

[Sagar-CK]

@DhavalRepo18 have updated the PR, tests still pass, let me know.