Add prompt subset support for benchmarks

[tlrmchlsmth]

Summary

This PR lets us limit the number of prompts we're working with if we want to jack up the prefix cache hit rate in evals or benchmarks

• add workload-level prompt_subset / prompt_subset_seed config and Starlark support
• add eval CLI flags for cycling a deterministic GSM8K/GPQA subset for duration-based runs
• partition prompt subsets globally before worker assignment to avoid worker collisions
• replay subsets by stable dataset indices/seeds/offsets instead of storing full prompt conversations in memory

Tests

• GOCACHE=/Users/tms/code/nyann-bench/.gocache go test ./... -count=1
• GOCACHE=/Users/tms/code/nyann-bench/.gocache go test ./pkg/dataset -count=1