improve Triton profiler (#191)

- add profiling of categories
- introduce more weights for relative runtime costs
- print relative time for all tasks and sub-tasks

The usual `prof_start!(…)` macro can be used as before. The same macro
now accepts an additional argument, a category: calling
`prof_start!(profiler, "task name", "category")`
will accumulate the times of all tasks with the same category and print
additional reports for the categories. For example:

```markdown
hash    84.60ms  8.88%
AIR     80.87ms  8.48%
cc     485.42ms 50.93%
LDE    165.70ms 17.38%
```