Ways the HUD compilers dashboard could be better

[zou3519]

I got here because I'm trying to answer the question of "which compiler benchmarks regressed in the past year?" I've spent a couple of hours on the HUD dashboard page, and I still haven't figured this out yet. Here's some of the gripes that I ran into while trying to answer this question.

1. The page seems to refresh itself every couple of minutes. This disrupts the train of thought. Also, I am not sure if the settings change when it refreshes.
2. The passrate chart and the graphs don't have all of the data. In particular, the passrate chart doesn't contain the max_autotune configs. I don't know how to actually click into the max_autotune data.

3. Wrong datapoints on PT2 dashboard for retried workflows test-infra#6173
4. There's one passrate chart but there are 3 passrate graphs. Scrolling between the graphs is kind of annoying
5. The graphs have so many series that some of them are hidden. Might be nicer to increase the height?

6. It's not clear to me how to hack on these charts. Using our internal tools (like scuba and unidash), it's easy (and well-known) on how to look up information.

Hypothesis: If we feed the data to internal sources and use internal tooling as the UXs, then we would be more productive than trying to roll our own UX.

cc @ZainRizvi @kit1980 @huydhn @clee2000

[huydhn]

cc @yangw-dev

(2) is the same as pytorch/test-infra#6175