feat(perf): add --apply-template flag to control genai prompt templating

[DingmaomaoBJTU]

Closes #1030 (follow-up to #1015).

Problem

winml perf --runtime winml-genai always wraps --prompt in the bundle's own chat template before timing. There was no way to benchmark a prompt that is already formatted (or a raw completion prompt), so users couldn't test their own templated input.

Change

Add --apply-template/--no-apply-template (default on, preserving current behavior). With --no-apply-template, the prompt is encoded and timed verbatim.

• GenaiPerfConfig.apply_template gates GenaiPerfBenchmark._prompt_text; when disabled, the prompt is used as-is with no apply_chat_template call.
• apply_template is recorded in the JSON report's benchmark_info for reproducibility.
• The flag is wired through the perf command and documented in --prompt / --apply-template help text.

Usage

# Default: prompt wrapped in the bundle's chat template
winml perf -m path/to/bundle --runtime winml-genai --prompt "Hello"

# Benchmark a pre-formatted prompt verbatim
winml perf -m path/to/bundle --runtime winml-genai \
  --no-apply-template --prompt "<|user|>Hello<|end|><|assistant|>"

[timenick]

I’m not sure why we need this user scenario, since onnxruntime-genai already includes an apply template API for it.

[DingmaomaoBJTU]

This is the scenario from #1030 (the issue this PR closes): benchmarking a prompt that is already formatted, instead of having perf wrap --prompt in the single-turn chat template on every run.

The default (--apply-template) still goes through ort-genai's apply_chat_template, so that path is unchanged. --no-apply-template is the opposite direction: it feeds the prompt to the tokenizer verbatim so you can time the exact token sequence you care about — a multi-turn conversation, a RAG prompt with a large retrieved context, or a custom/non-default template. Since perf measures prefill (TTFT) and decode throughput, the prompt's real length and shape drive the numbers; the single-turn template wrapper doesn't represent those cases. The genai apply_chat_template API templates a prompt, but doesn't give perf a way to skip templating and benchmark a pre-formatted one, which is what this flag adds.