Improve request-level data collection when only the metrics are needed.

[azamikram]

What would you like to be added:

A new flag, such as metrics_only, designed to work in conjunction with per_request. When enabled, this flag would restrict data collection exclusively to computable performance metrics, such as TTFT, TPOT, and ITLs.

Why is this needed:

When running inference-perf with high QPS and the per_request flag enabled, generating and storing the final per_request_lifecycle_metrics.json file takes a significant amount of time. This delay occurs because the log captures raw request and response content for every individual request, and hence making disk I/O a major bottleneck.

In high-QPS cases, this file can grow to several gigabytes, delaying report generation times to tens of minutes. This overhead is unnecessary for users who do not need the raw request/response payloads and are only interested in analyzing the per-request timing metrics (TTFT, TPOT, and ITL).

[jjk-g]

Feature request is to ensure latency metrics are available in per_request

More holistically it might make sense to control some more of the per_request attributes with config bools to make it more granular.