Increase several test tolerances to ensure tests pass on more computers

[martinholmer]

Even after the substantial reweighting improvements in PR #407, the reweighting results are not exactly the same on all computers. The changes in this pull request increase some of the numpy.allclose tolerances so that the tests also pass when the reweighting is done on Apple M4 CPU chips.

[donboyd5]

Not surprisingly, it passes all tests on my gpu. I will not be able to run it on my cpu until tomorrow. However, I neglected to run test on my cpu before doing my final pr. I got the following errors last night. While it's not a problem for me because I will use gpu, it suggests that others with different cpu setups might get slightly different results. It might make sense to expand bounds slightly for these variables. I hope to have a chance to look for ways to tighten up cross-machine differences further but it will be awhile before I can do that.

I can send cpu results tomorrow if you want.

Personally I think it would be fine to merge as or with additional tolerance increases to deal with the cases below.

E           WEIGHT_DIFF:sdev,act,exp= 961.6711634863782 961.7270821801824; **needs expansion**

E           ! YR,KIND,EST= 2023 paytax 1381.9 actual; ok within .01
E           ! YR,KIND,EST= 2023 paytax 1381.8 expected

E           ! YR,KIND,EST= 2023 qbid 52.7 actual; **needs .01**
E           ! YR,KIND,EST= 2023 qbid 52.6 expected

E           DIFF:OTM,totben,act,exp= 23.94 23.95 ok within .01
E           DIFF:OTM,affpct,act,exp= 8.89 8.9 **needs .01**
E           DIFF:ALL,totben,act,exp= 54.85 54.86 needs **.01**

[donboyd5]

@martinholmer, I ended up getting a chance to run it cpu-only on my machine. I had one test failure, below. It looks like loosening the tolerance on OTM,affpct to .01 would get all to pass on my machine cpu-only. Eventually, if others run it we may find additional small differences. But maybe by then I'll have a PR to get it to solve with more precision in reasonable time, so that looser tolerances won't be necessary.

E           ValueError: 
E           IMPUTED VARIABLE DEDUCTION BENEFIT ACT-vs-EXP DIFFS:
E           DIFF:OTM,affpct,act,exp,atol= 8.89 8.9 1e-08

[martinholmer]

@donboyd5 reported these seemingly contradictory CPU-based test differences an hour apart:

E           WEIGHT_DIFF:sdev,act,exp= 961.6711634863782 961.7270821801824; **needs expansion**

E           ! YR,KIND,EST= 2023 paytax 1381.9 actual; ok within .01
E           ! YR,KIND,EST= 2023 paytax 1381.8 expected

E           ! YR,KIND,EST= 2023 qbid 52.7 actual; **needs .01**
E           ! YR,KIND,EST= 2023 qbid 52.6 expected

E           DIFF:OTM,totben,act,exp= 23.94 23.95 ok within .01
E           DIFF:OTM,affpct,act,exp= 8.89 8.9 **needs .01**
E           DIFF:ALL,totben,act,exp= 54.85 54.86 needs **.01**

and then

E           ValueError: 
E           IMPUTED VARIABLE DEDUCTION BENEFIT ACT-vs-EXP DIFFS:
E           DIFF:OTM,affpct,act,exp,atol= 8.89 8.9 1e-08

@donboyd5, I completely confused. Why the difference in these two sets of results?

[donboyd5]

@martinholmer, sorry for the confusion. The 2nd one is the correct one.

After updating to commit 0f9dc5d all non-skipped tests pass when running on cpu-only.