feat(M100): Request Payload Size Impact Analysis

[hlin99]

Problem

Users need to understand how prompt size affects inference performance to optimize their workloads.

Solution

Add xpyd-bench size-impact CLI subcommand that sweeps across prompt sizes and measures performance at each level.

Acceptance Criteria

• xpyd-bench size-impact --base-url <url> --model <model> CLI subcommand
• Sweep across prompt sizes (default: 10, 100, 500, 1000, 2000, 4000 tokens)
• Measure TTFT, TPOT, throughput at each prompt size level
• Detect linear vs sublinear vs superlinear scaling behavior
• Report: size-latency curve, inflection points, recommended max prompt size for target latency
• JSON output with per-size-level results
• --size-levels to customize sweep sizes
• Tests covering sweep orchestration, scaling detection, and CLI integration