bench: cycle distinct prefilled blocks to remove per-append fill

[nclack]

Closes acquire-project#150.

Streaming benches spent ~half their wall time generating their own input via the per-append fill() in pump_data / pump_data_interleaved, which understated library throughput ~2x and let the fill thread contend with the library for host memory bandwidth (medfmt/zstd regressed ~18% purely from contention).

Changes:

• New default pump mode cycles through up to 8 distinct pre-generated blocks (PUMP_CYCLE_BLOCK_COUNT), filled once with staggered offsets so contents differ block-to-block but keep the same value distribution — realistic compression ratios with zero per-append fill cost. Applied to both single-stream and the two-stream interleaved pump.
• --busy-producer keeps the original regenerate-every-append behavior as a host-bandwidth stress variant.
• --prefill (single identical block) retained for back-compat / writer-only ceiling.
• Both pumps now report Data gen time (seconds / % of wall) so harness cost is visible instead of folded into throughput.
• Incidental: the single-stream path previously ignored --dtype (always bpe=2); it now uses dtype_bpe(dtype), matching the two-stream path.

A/B to validate: default vs --busy-producer throughput (expect ~2x gap; medfmt/zstd contention regression gone under default), and matching compressed bytes between default and busy modes (vs inflated ratio under --prefill).

[nclack]

Superseded by acquire-project#159 (PR retargeted to upstream).