Add MD.Clone function

[djdongjin]

MD is not thread-safe so sometimes we need to make a copy before modification, e.g. in containerd,

https://github.com/containerd/containerd/blob/c3efa0cb339ffc2d6c9f03d7fb4e5e6577d0194b/pkg/namespaces/ttrpc.go#L30-L36

This PR adds a MD.Clone function, similar to golang's http.Header.Clone (the code is copied there). Also add a benchmark that shows golang's implementation is better than a simple copy approach (both memory&latency):

$ go test -bench='Benchmark.*Clone' -benchmem
goos: linux
goarch: amd64
pkg: github.com/containerd/ttrpc
cpu: Intel(R) Xeon(R) Platinum 8259CL CPU @ 2.50GHz
BenchmarkMetadataClone/size=5-4          2663775               494.2 ns/op            80 B/op          1 allocs/op
BenchmarkMetadataClone/size=10-4          707372              1608 ns/op             871 B/op          2 allocs/op
BenchmarkMetadataClone/size=20-4          371546              2829 ns/op            1785 B/op          2 allocs/op
BenchmarkMetadataClone/size=50-4          194526              5800 ns/op            4436 B/op          3 allocs/op
BenchmarkSimpleMetadataClone/size=5-4            1809780               679.0 ns/op            80 B/op          5 allocs/op
BenchmarkSimpleMetadataClone/size=10-4            525523              2128 ns/op             871 B/op         11 allocs/op
BenchmarkSimpleMetadataClone/size=20-4            272677              4355 ns/op            2583 B/op         22 allocs/op
BenchmarkSimpleMetadataClone/size=50-4            100525             10794 ns/op            6798 B/op         55 allocs/op
PASS
ok      github.com/containerd/ttrpc     11.859s

[klihub]

nit: I think we shouldn't copy this comment verbatim from the HTTP implementation. Any mention of a ReverseProxy is a potential source for confusion, as ttrpc which is typically used for communication between two process on the same physical machine, using a unix domain socket as the transport.

Maybe we should just leave

if vv == nil { // Preserve nil values.

here.

[djdongjin]

Thanks for the suggestion @klihub . SGTM, updated the PR.

[djdongjin]

Hi @fuweid thanks for approving/merging this PR! I wonder if we could do a ttrpc release recently (ideally also include this dependency upgrade PR #178 but it's optional).

This is mainly to fix a concurrent map write panic in otelttrpc (containerd/otelttrpc#2), which is reported in containerd/containerd#11138)

thanks

[fuweid]

ping @djdongjin https://github.com/containerd/ttrpc/releases/tag/v1.2.7 this is new tag.

[djdongjin]

ping @djdongjin https://github.com/containerd/ttrpc/releases/tag/v1.2.7 this is new tag.

Great, thanks for the quick release!