Reduce memory allocations when rendering diff flamegraphs

[aleks-p]

When we render large diff flamegraphs, we create quite a bit of allocations due to prepending to a slice instead of appending. We also use an older and less efficient way of truncating the diff tree, I've replaced that with the one used for truncating regular (non-diff) trees

goos: darwin
goarch: arm64
pkg: github.com/grafana/pyroscope/pkg/model
cpu: Apple M2 Pro
                      │ reference-diff.txt │         optimized-diff.txt          │
                      │       sec/op       │   sec/op     vs base                │
_NewFlamegraphDiff-10         3956.7m ± 3%   252.0m ± 7%  -93.63% (p=0.000 n=10)

                      │ reference-diff.txt │          optimized-diff.txt          │
                      │        B/op        │     B/op      vs base                │
_NewFlamegraphDiff-10       18313.8Mi ± 0%   749.1Mi ± 0%  -95.91% (p=0.000 n=10)

                      │ reference-diff.txt │         optimized-diff.txt          │
                      │     allocs/op      │  allocs/op   vs base                │
_NewFlamegraphDiff-10          1.489M ± 0%   1.053M ± 0%  -29.33% (p=0.000 n=10)

[aleks-p]

this is based on model.Tree.minValue(), adapted for 2 trees

[simonswine]

LGTM

Thanks for spotting this. Wonder how "good" your benchmarks were, but might make sense to add them.

[aleks-p]

Wonder how "good" your benchmarks were, but might make sense to add them.

I can't add the data I used for the benchmark, but I can likely capture the improvement with synthetic data as well. I will add the benchmark code before merging.

Edit: done