Refactor and update benchmarks

[charleskawczynski]

This PR updates and refactors the benchmarks in a few ways:

• If a column is the same for all benchmarks, then simply include that information at the table's header and remove the column
• Starts a dictionary for device specs and uses CUDA.name(CUDA.device()) to look up the device specs

function device_info()
    device_specs =
        Dict("NVIDIA A100-SXM4-80GB" => (; device_bandwidth_GBs = 2_039))
    is_cuda = ClimaComms.device() isa ClimaComms.CUDADevice
    if is_cuda && haskey(device_specs, CUDA.name(CUDA.device()))
        (; device_bandwidth_GBs) = device_specs[CUDA.name(CUDA.device())]
        return (;
            device_bandwidth_GBs,
            exists = true,
            name = CUDA.name(CUDA.device()),
        )
    else
        return (;
            device_bandwidth_GBs = 1,
            exists = false,
            name = CUDA.name(CUDA.device()),
        )
    end
end

Now, we can add more devices in device_specs.

• Pass in kernel_time_s to push_info so that it's decoupled from nreps. This is clearer and becomes compatible with benchmarks generated from BenchmarkTools
• Pass in problem_size into push_info, so that a benchmark table can have multiple problem sizes (this is helpful for our copyto benchmarks)
• Use the benchmark utils in the copyto! benchmarks.