Add Anthropic Messages API support

inference-perf supported OpenAI-style completion and chat APIs, but not Anthropic's /v1/messages contract. That limited benchmarking fidelity for Claude-style agentic workloads, tool-use round trips, and OTel trace replay through Anthropic-compatible serving stacks.

Implemented an Anthropic Messages API adapter with request serialization, system prompt and tool schema conversion, tool_use and tool_result mapping, non-streaming and streaming response parsing, client/datagen wiring, docs, CLI flag updates, and regression coverage.

• Extends inference-perf to benchmark Anthropic-compatible endpoints alongside OpenAI-compatible endpoints.
• Improves replay fidelity for agentic traces that naturally use Anthropic message, tool-use, and tool-result semantics.
• Adds reusable support across mock data generation, ShareGPT conversion, vLLM/mock clients, and OTel replay paths.

• Added request-formatting, response-parsing, config, client-header, streaming, and replay tool-use capture coverage.
• Addressed reviewer feedback around streaming behavior and tool-call coverage before approval.
• PR validation covered ruff format/check, strict mypy, focused pytest suites, OTel replay tests, and CLI docs sync.