Add support for fixed schedule warmup

[matthewkotila]

Previously, someone would run PA fixed schedule mode like this:

perf_analyzer \
  --fixed-schedule \
  --input-data=input_data.json \
  -m facebook/opt-125m \
  --service-kind=openai \
  --endpoint=v1/chat/completions \
  --async

with an input_data.json like this:

{
  "data": [
    {
      "payload": [{
          "model": "facebook/opt-125m",
          "messages": [{"role": "user","content": "my_prompt_1"}],
          "max_completion_tokens": 1
        }],
      "timestamp": [1000]
    },
    {
      "payload": [{
          "model": "facebook/opt-125m",
          "messages": [{"role": "user","content": "my_prompt_2"}],
          "max_completion_tokens": 1
        }],
      "timestamp": [2000]
    }
  ]
}

But they wouldn't be able to use the warmup feature (--warmup-request-count).

Now, with this PR, users can use the warmup feature with the fixed schedule feature:

perf_analyzer \
  --fixed-schedule \
  --warmup-request-count=1 \
  --input-data=input_data.json \
  -m facebook/opt-125m \
  --service-kind=openai \
  --endpoint=v1/chat/completions \
  --async

Basically, if --warmup-request-count=N, the first N payloads in input_data.json are sent as "warmup" requests (i.e. excluded from the final performance metric/statistic calculations and profile export JSON), and the rest are part of the standard benchmark.

[Copilot]

Pull Request Overview

This PR adds support for fixed schedule warmup by modifying request scheduling, adjusting input parsing, and updating command-line validations. Key changes include:

• Introducing a new dataset_offset parameter in worker and thread configuration classes.
• Modifying the ModelParser to initialise fixed schedule inputs via a new constructor.
• Refactoring CustomRequestScheduleManager to distinguish between warmup and benchmark schedules.

Reviewed Changes

Copilot reviewed 25 out of 25 changed files in this pull request and generated no comments.

Show a summary per file

File	Description
src/request_rate_worker.h	Added dataset_offset parameter to worker construction.
src/request_rate_manager.h & .cc	Updated thread configuration and worker creation APIs.
src/perf_analyzer.cc	Adjusted ModelParser instantiation for fixed schedule.
src/model_parser.{h,cc}	Introduced a constructor that initializes fixed schedule inputs and removed duplicate code in InitOpenAI.
src/custom_request_schedule_manager.{h,cc}	Refactored schedule generation to separate warmup and benchmark schedules.
src/command_line_parser.cc	Updated help messages and validation checks for fixed schedule mode.
Test and documentation files	Updated to support new warmup functionality.

[nicolasnoble]

Yeah looks good to me. The change is large due to propagation of some of the parameters, but that's mostly mechanical work. I've left only a few nits here and there, otherwise it's fine with me.