Revert "[Utilties] don't throw UnsupportedAttribute in pass_attribute…

[odow]

This reverts commit 1a3c637.

This caused widespread failures in SolverTests for solvers using a caching optimizer.
#1656 (comment)

[blegat]

When you have a CachingOptimizer in EMPTY_OPTIMIZER state, I think the behavior of setting an attribute unsupported by the optimize or copy_toing a model with an attribute unsupported by the optimizer should be the same. With this PR, it will throw on the copy but not on set. Adding the check in

MathOptInterface.jl/src/Utilities/cachingoptimizer.jl

Line 729 in b552036

if m.state == ATTACHED_OPTIMIZER

would throw it in both cases.
Note that if we do it for the attributes, we should do it for add_constraint too.

Another (simpler IMO) solution would be to modify the test into something like

@test_throws ... begin
    copy_to(...)
    optimize!(...)
end

[odow]

Yeah okay. Maybe it's a bad test. We don't have a lot of consistency around when errors are thrown.

[odow]

Closing in favor of #1686