Add two-run pattern for video_transcode_bench timed mini to reduce runtime

[marziehlenjaniMeta]

Summary:
The video_transcode_bench_svt_timed_mini benchmark takes ~36 seconds
wall-clock, but only ~15 seconds is actual encoding. The remaining ~19
seconds is spent downscaling all 42 source clips to target resolutions — a
data preparation step, not representative of production workloads.

This diff introduces a two-run pattern (similar to DjangoBench and
SparkBench mini) that separates preparation from measurement:

• video_transcode_bench_svt_timed_mini_prep: Downscales a sampled subset
  of clips (20%, ~8 clips) and caches them in resized_clips/ for reuse.
  After this run, the ~42GB source dataset can be deleted to reduce OS image
  size.
• video_transcode_bench_svt_timed_mini_reuse: Skips downscaling entirely,
  reuses cached resized_clips/, and runs only the encoding workload. Errors
  if the cache doesn't exist. Can be run repeatedly.

Additional changes:

• Split the generated run script so downscaling runs during preprocessing
  (tracked as a downscaling sub-operation in breakdown.csv), and only
  encoding runs during main_benchmark
• Replaced the sleep+SIGTERM time-limiting mechanism in
  timed_parallel_feeder.sh with a deadline-based feeding loop that stops
  dispatching jobs after max_time
• Reduced max_time from 15 to 10 seconds for both new variants
• Added --skip-downscale and --keep-downscaled flags to run.sh
• Updated README with two-run pattern instructions

Differential Revision: D102686096

[meta-codesync]

@marziehlenjaniMeta has exported this pull request. If you are a Meta employee, you can view the originating Diff in D102686096.