cue: Go API downloads more than necessary for simple use cases

[mvdan]

Right now, a simple example like https://pkg.go.dev/cuelang.org/[email protected]/cue#example-Context downloads quite a lot of modules:

go: finding module for package cuelang.org/go/cue
go: finding module for package cuelang.org/go/cue/cuecontext
go: downloading cuelang.org/go v0.4.3
go: downloading github.com/cockroachdb/apd/v2 v2.0.1
go: downloading github.com/cockroachdb/apd v1.1.0
go: downloading github.com/mpvl/unique v0.0.0-20150818121801-cbe035fff7de
go: downloading golang.org/x/text v0.3.7
go: downloading golang.org/x/net v0.0.0-20200226121028-0de0cce0169b
go: downloading github.com/google/uuid v1.2.0
go: downloading github.com/pkg/errors v0.8.1
go: downloading gopkg.in/yaml.v3 v3.0.0-20210107192922-496545a6307b

Some of these look unnecessary, like pkg/errors and apd v1. We should trim these down.

[mvdan]

We can drop pkg/errors once cockroachdb/apd#123 is done.

[mvdan]

Upstream apd fixed their reliance on pkg/errors, though they did that in v3, while we are in v2. @mpvl @myitcv would it be OK to upgrade to apd/v3 now? Per https://github.com/cockroachdb/apd/releases/tag/v3.0.0 there's only one breaking change, so it shouldn't be a hard upgrade. We probably also don't want to stay on v2 for too long, as it hasn't seen an update in over two years now.

[mvdan]

I gave it a quick go, and it mostly works, though a few tests are unhappy due to different resulting precision. For example, in master with a bit of debug logging to see the apd.Decimal being marshaled:

$ go test -v -run TestMarshalJSON/22 ./cue
=== RUN   TestMarshalJSON
=== RUN   TestMarshalJSON/22/a:_1.0/1
apd.Decimal{Form:0, Negative:false, Exponent:-1, Coeff:big.Int{neg:false, abs:big.nat{0xa}}}
"1.0"
--- PASS: TestMarshalJSON (0.00s)
    --- PASS: TestMarshalJSON/22/a:_1.0/1 (0.00s)
PASS
ok  	cuelang.org/go/cue	0.005s

and now in my branch:

$ go test -v -run TestMarshalJSON/22 ./cue
=== RUN   TestMarshalJSON
=== RUN   TestMarshalJSON/22/a:_1.0/1
apd.Decimal{Form:0, Negative:false, Exponent:-23, Coeff:apd.BigInt{_inner:(*big.Int)(nil), _inline:[2]big.Word{0x2c7e14af6800000, 0x152d}}}
"1.00000000000000000000000"
    types_test.go:2721: 
         got {"a":1.00000000000000000000000};
        want {"a":1.0}
--- FAIL: TestMarshalJSON (0.00s)
    --- FAIL: TestMarshalJSON/22/a:_1.0/1 (0.00s)
FAIL
FAIL	cuelang.org/go/cue	0.006s

This seems like a change that we don't want, because we add verbosity for no reason. However I can't tell if it's a bug on our side (perhaps relying on undocumented behavior), or an intentional change in behavior in upstream's v3, which they are allowed to do per semver. I could not find anything in their release changelogs.

Worth noting that I think the new Exponent:-23 comes from CUE using an apd precision of 24.

I think upstream changed this in cockroachdb/apd#115, as it removes calls to Reduce from funcs like Quo, which is about removing trailing zeros. And this test in particular performs a division.

[mvdan]

I should have clarified that the CL is https://review.gerrithub.io/c/cue-lang/cue/+/543421, and the full results are at https://github.com/cue-lang/cue-trybot/actions/runs/3045646884/jobs/4907455935.