Reorganize the turnkey benchmark help page

[jeremyfowers]

Problem Statement

The current help page for turnkey benchmark -h presents a lot of options in no particular order. This doesn't give new users a good path to parsing the available options and associating them with their respective phases of the toolchain.

Proposed Solution

Reorganize the help page like this:
 

Discover, build, and then benchmark the model(s) within input file(s).

positional arguments:
  input_files           One or more script (.py), ONNX (.onnx), or input list (.txt) files to be benchmarked

options:
  -h, --help            show this help message and exit

options that specifically apply to the `discover` part of the toolflow:
  --script-args SCRIPT_ARGS
                        Arguments to pass into the target script(s)
  --max-depth MAX_DEPTH
                        Maximum depth to analyze within the model structure of the target script(s)
  --labels [LABELS [LABELS ...]]
                        Only evaluate the scripts that have the provided labels

options that apply to both the `build` and `benchmark` parts of the toolflow:
  --device {nvidia,x86}
                        Type of hardware device to be used for the benchmark (defaults to "x86")
  --runtime {ort,trt,torch-eager,torch-compiled}
                        Software runtime that will be used to collect the benchmark. Must be compatible with the selected device. Automatically selects a sequence if `--sequence` is not used.If this argument is not set, the default runtime of the selected device will be used.
  -d CACHE_DIR, --cache-dir CACHE_DIR
                        Build cache directory where the resulting build and benchmarking artifacts will be stored (defaults to /home/jfowers/.cache/turnkey)
  --lean-cache          Delete all build artifacts except for log files when the command completes

options that apply specifically to the `build` part of the toolflow:
  --sequence {optimize-fp16,optimize-fp32,onnx-fp32}
                        Name of a build sequence that will define the model-to-model transformations, used to build the models. Each runtime has a default sequence that it uses.
  --rebuild {if_needed,always,never}
                        Sets the cache rebuild policy (defaults to if_needed)
  --onnx-opset ONNX_OPSET
                        ONNX opset used when creating ONNX files (default=14). Not applicable when input model is already a .onnx file. options that apply specifically to the `benchmark` part of the toolflow:

options that apply specifically to the `benchmark` part of the toolflow:
  --iterations ITERATIONS
                        Number of execution iterations of the model to capture the benchmarking performance (e.g., mean latency)
  --rt-args [RT_ARGS [RT_ARGS ...]]
                        Optional arguments provided to the runtime being used

options that apply to all toolflows:
  --use-slurm           Execute on Slurm instead of using local compute resources
  --process-isolation   Isolate evaluating each input into a separate process
  --timeout TIMEOUT     Build timeout, in seconds, after which a build will be canceled (default=3600). Only applies when --process-isolation or --use-slurm is also used.