[Presets] Make a preset workload state the requested dataset

[peterschmidt85]

Fixes #4198.

random is dstack's name for a synthetic workload; a benchmark tool has its own name for the data
it generates — random in vLLM, generated-shared-prefix in SGLang. Verification compared the
two as one name, so a preset with shared_prefix_tokens and no dataset failed whenever the
agent benchmarked with SGLang, and passed with vLLM only because the names coincide.

Now dataset states the dataset you asked for, and nothing else.

type: preset
base: Qwen/Qwen3.8-27B

input_tokens: 131072
output_tokens: 512
shared_prefix_tokens: 130048

No dataset, so the benchmark runs synthetic prompts. Verification compares the shared prefix and
the concurrency with the request; whatever the tool calls its generated data stays in
benchmark.command, and the stored preset records no dataset.

type: preset
base: Qwen/Qwen3.8-27B

dataset: spec_bench

dataset names a real dataset, which defines its own requests. Verification checks the benchmark
served exactly it, and every mismatch names both values.

The UI shows a preset's dataset only when one was requested.

Breaking changes

• dataset: random is rejected: omit dataset for synthetic prompts.

Backward compatibility

• Presets and in-flight sessions that recorded dataset: random upgrade when read.

Reworks #4199 by @r4victor; its verification comparisons and much of its test suite carry over.
Reworked because:

1. Fix preset verification rejecting a tool's synthetic dataset name #4199 kept the tool's name in workload.dataset, and the pulled-preset row from [Presets] Support for dstack preset push and dstack preset pull #4194 reads
   that field as the request — the same confusion, one layer up.
2. A field that echoes the request for a named dataset but records a tool detail for a synthetic
   one has two meanings; now it has one.
3. The contract holds without system prompt changes: the stored workload is normalized from the
   configuration.

AI assistance: written with Claude Code.