Add fields for working GPU count in rollup data and use working GPU count to compute total_throughput_per_gpu

[HuiGao-NV]

We may allocate part of devices on one node, such as 1 prefill server with DEP4 only needs 4 GPUs.
With current formula, this server will be counted with 8 GPUs although it only use 4.
The total throughput per gpu is more accurate to be computed with the real number of working GPUs, so add new logic to get the number of real working GPU devices.

[codecov-commenter]

Codecov Report

✅ All modified and coverable lines are covered by tests.
⚠️ Please upload report for BASE (main@c180328). Learn more about missing BASE report.

Additional details and impacted files

@@           Coverage Diff           @@
##             main     #270   +/-   ##
=======================================
  Coverage        ?   67.99%           
=======================================
  Files           ?       70           
  Lines           ?     9304           
  Branches        ?        0           
=======================================
  Hits            ?     6326           
  Misses          ?     2978           
  Partials        ?        0           

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[HuiGao-NV]

669968.tar.gz
This is the raw data of one slurm job.
The changed data fields are in 669968/logs/benchmark-rollup.csv:

Config,Total GPU Count,Decode GPU Count,Total Working GPU Count,Decode Working GPU Count,Prefill Working GPU Count,Concurrency,Total Token Throughput,Output Token Throughput,Median TTFT,Median TPOT,Median ITL,P90 Decode Running Requests,Output Token Throughput per User,Total Token Throughput per GPU
disagg-B300-1p5d-dep4-tep4-c195-stp,32,24,24,20,4,195,86848.766,9631.879,3286.376,15.534,1548.842,,64.377,3618.699

We can see that the "Total GPU Count" is 32 while the "Total Working GPU Count" is 24 with 4 prefill GPUs sand 20 decode GPUs. This is consistent with config in 669968/config.yaml:

resources:
gpu_type: B300
gpus_per_node: 8
prefill_nodes: 1
prefill_workers: 1
gpus_per_prefill: 4
decode_nodes: 3
decode_workers: 5
gpus_per_decode: 4