bench: add direct compiled timing

[karthickai]

Adds default-enabled direct compiled timing as compiled_nocudagraphs_us while preserving existing cudagraph timing in compiled_us.

Use --no-compiled-nocudagraphs to disable it.

[eellison]

Can we clarify whether this metric is intended to measure default Inductor without CUDAGraphs or coordinate descent without CUDAGraphs? The name/docs imply default, but because this runs after compiled_cd, it currently measures CD code

[karthickai]

Can we clarify whether this metric is intended to measure default Inductor without CUDAGraphs or coordinate descent without CUDAGraphs? The name/docs imply default, but because this runs after compiled_cd, it currently measures CD code

thanks for catch, the intended measures default Inductor without CUDAGraphs. I moved the timing before CD = so it uses the original default compiled artifact.