Download benchmark results: set concurrent range gets to 8

[assafvayner]

From the results of the benchmarks, we should bump up the number of ranges downloaded at a time to a maximum of 8. With no change to the number of threads used.

---

details:

• the benchmark is not comprehensive over all data types. Only file tested was a 5 GB safetensors file with 75 terms and largely no dedupe. The environment used was a t3.2xl ec2 instance; fast link to s3.
• the benchmark showed that the current chunk cache is a significant bottleneck (causing download times to nearly triple ~15s -> ~40s). Cache changes will be addressed in later pull request/s.
• Changing the number of system threads used had largely no/minor impact on the download time.
• Changing the number of maximum buffered terms did have an impact.
  • with caching disabled 16 was marginally better than 8
  • with caching enabled 8 was marginally better than 16
  • in any case 4 (current) was too low
  • attempting to download all terms at once has a marginal benefit with caching disabled, while yielding worse performance with the current caching implementation likely stemming from lock contention.