feat: optimized pprof symbolication

[kolesnikovae]

The main goal of the change is to make generated pprof proto type the only profile representation we deal with in the backend:

• Use of profile.Profile does not allow implementing node truncation.
• Tree representation we currently use for building a flamegraph does not allow including location information. Synthetic benchmarks show that use of pprof proto Profile decreases CPU time needed to built a profile by up to 300% (1.5s -> 350ms), and allocated memory by up to 100%.

Next steps:

1. Add support for max_nodes parameter to pprof query #2728
2. Profile truncation should preserve top-N stack traces with largest "self"  #2729
3. Switch to pprof in SelectMergeStacktraces API

Dependent issues:

1. Add locations to the flamegraph #2727
2. Function metric #2228

[knylander-grafana]

@kolesnikovae This looks like an interesting feature. Should we add doc for this?

[cyriltovena]

3. Switch to pprof in SelectMergeStacktraces API

This is more an performance optimisation so far so not much to document yet.

[cyriltovena]

LGTM

[kolesnikovae]

@kolesnikovae This looks like an interesting feature. Should we add doc for this?

The change only affects internals with no impact on the end user in terms of the system behaviour and mechanics. This is a preparation step for some of the features we want to implement (for which we will need docs, indeed)