[Analytics] Add token throughput metrics (tokens/second)

[PierreLeGuen]

Description

Track and expose token throughput metrics to measure inference performance.

Motivation

Tokens per second is a key performance metric for LLM inference. This helps identify:

• Model performance characteristics
• Infrastructure bottlenecks
• Capacity planning needs

Proposed Solution

Add throughput metrics:

• Tokens per second (output tokens / decoding time)
• Average throughput by model
• P50/P95/P99 throughput percentiles

Technical Details

• Calculate from existing latency data: output_tokens / decoding_time_ms * 1000
• Aggregate in analytics queries
• Consider storing as pre-computed metric

Acceptance Criteria

• Throughput metric available in API responses
• Breakdown by model
• Percentile distributions (P50, P95, P99)