Optimize repeated row iterator

[kolesnikovae]

The PR is aimed at improving performance of the parquet rows iterator:

1. Refactored repeated row iterator, fixed excessive CPU time consumption.
2. Asynchronous batching of fetched rows.
3. Iterator tee. Instead of copying row iterators into memory, we now can stream them.

Benchmarking on synthetic data does not show any substantial improvements. However, with the real life data and queries, the difference is quite prominent:

goos: darwin
goarch: arm64
pkg: github.com/grafana/pyroscope/pkg/phlaredb
                             │  before.txt  │              after.txt              │
                             │    sec/op    │   sec/op     vs base                │
_MergeSampleByStacktraces-10   1278.8m ± 6%   463.1m ± 2%  -63.79% (p=0.000 n=10)

                             │  before.txt  │              after.txt              │
                             │     B/op     │     B/op      vs base               │
_MergeSampleByStacktraces-10   318.4Mi ± 2%   307.2Mi ± 2%  -3.52% (p=0.000 n=10)

                             │  before.txt  │              after.txt              │
                             │  allocs/op   │  allocs/op   vs base                │
_MergeSampleByStacktraces-10   1193.3k ± 0%   884.8k ± 0%  -25.85% (p=0.000 n=10)

Flamegraph

In terms of query latency, in some cases it is reduced by up to 25-30%, with the impact becoming more noticeable the more profiles that need to be retrieved.

[cyriltovena]

100%

[cyriltovena]

Stellar !!!! This is going to be perfect for the future.

LGTM