[Lang] [test] Add ti.get_rel_eps() for maximal relative error tolerance on current backend

[k-ye]

It turns out that #1779 doesn't play well on my device. Adding this to make the allclose tests pass again on Metal.

Related issue = #1779

[Click here for the format server]

---

[codecov]

Codecov Report

Merging #1798 into master will not change coverage.
The diff coverage is n/a.

@@           Coverage Diff           @@
##           master    #1798   +/-   ##
=======================================
  Coverage   61.74%   61.74%           
=======================================
  Files          19       19           
  Lines        3850     3850           
  Branches      691      691           
=======================================
  Hits         2377     2377           
  Misses       1348     1348           
  Partials      125      125           

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update da41c85...a443e34. Read the comment docs.

[archibate]

LGTM now. We may also find a proper ti.get_abs_eps for each backend.

[k-ye]

We may also find a proper ti.get_abs_eps for each backend.

Actually, what's the difference between rel and abs in this case? I think i get it now after looking at the tests..

[archibate]

We may also find a proper ti.get_abs_eps for each backend.

Actually, what's the difference between rel and abs in this case?

IIUC it's roughly:

def erf(x):
  return max(abs_eps, rel_eps * x)