Convert Attribute to a struct

[semistrict]

This avoid about 30% of allocations in benchmarks with 3 or more
attributes.

name                     old time/op    new time/op    delta
StartEndSpan-8              522ns ± 0%     502ns ± 0%   ~     (p=1.000 n=1+1)
SpanWithAnnotations_3-8    1.22µs ± 0%    1.08µs ± 0%   ~     (p=1.000 n=1+1)
SpanWithAnnotations_6-8    1.66µs ± 0%    1.29µs ± 0%   ~     (p=1.000 n=1+1)
TraceID_DotString-8         191ns ± 0%     185ns ± 0%   ~     (p=1.000 n=1+1)
SpanID_DotString-8          172ns ± 0%     174ns ± 0%   ~     (p=1.000 n=1+1)

name                     old alloc/op   new alloc/op   delta
StartEndSpan-8               576B ± 0%      576B ± 0%   ~     (all equal)
SpanWithAnnotations_3-8    1.37kB ± 0%    1.27kB ± 0%   ~     (p=1.000 n=1+1)
SpanWithAnnotations_6-8    1.49kB ± 0%    1.30kB ± 0%   ~     (p=1.000 n=1+1)
TraceID_DotString-8         80.0B ± 0%     80.0B ± 0%   ~     (all equal)
SpanID_DotString-8          56.0B ± 0%     56.0B ± 0%   ~     (all equal)

name                     old allocs/op  new allocs/op  delta
StartEndSpan-8               4.00 ± 0%      4.00 ± 0%   ~     (all equal)
SpanWithAnnotations_3-8      13.0 ± 0%      10.0 ± 0%   ~     (p=1.000 n=1+1)
SpanWithAnnotations_6-8      18.0 ± 0%      12.0 ± 0%   ~     (p=1.000 n=1+1)
TraceID_DotString-8          3.00 ± 0%      3.00 ± 0%   ~     (all equal)
SpanID_DotString-8           3.00 ± 0%      3.00 ± 0%   ~     (all equal)

[semistrict]

Not sure why the delta column displays just "~" but you can see that the numbers did change. This test compares the numbers after making everything sampled (before, we didn't benchmark the sampled case, I also added sampled/not-sampled benchmarks).

[odeke-em]

Not sure why the delta column displays just "~" but you can see that the numbers did change

If you didn't already, could you please run the benchmarks say 10 times or so e.g

go test -run=^$ -bench=. -count=10

[semistrict]

Unfortunately it isn't easy to reproduce the above table with "count=10" because I changed the benchmarks themselves in this PR. In any case, the performance gain is not the only reason for this change - it is more idiomatic (returning structs rather than interfaces) and it's a net code reduction (except for more test code).