[PERF] Add a parallel local CSV reader

[desmondcheongzx]

Adds a parallel CSV reader to speed up ingestion of CSV. The approach adapts some ideas laid out in [1], but the majority of performance gains came from the use of buffer pools to minimize memory allocations.

Some performance numbers

We consider a simple case of reading and performing .collect() on a CSV file with 10^8 rows of 9 fields: 3 string fields, 5 int64 fields, and 1 double field. This file is roughly 5GB in size.

Non-native executor:                 38.71140212500177s
Non-native executor, new CSV reader: 7.432862582994858s
Native executor:                     44.55550079200475s
Native executor, new CSV reader:     4.117344291880727s

This represents a roughly 10x speedup on CSV reads for the native executor.

Followup work

• The schema is currently taken from the convert options, but we should also perform schema inference.
• We need to add better estimators for record size, either via sampling, or by keeping track of stats as we go.
• Currently, for each read, the reader creates a buffer pool for reading CSV records plus a pool of slabs for reading the CSV file. We might need to change these to per-process pools to avoid high memory pressure on concurrent reads. However care must be taken otherwise it's possible for use to deadlock.

[1]: Ge, Chang et al. “Speculative Distributed CSV Data Parsing for Big Data Analytics.” Proceedings of the 2019 International Conference on Management of Data (2019).

[codspeed-hq[bot]]

CodSpeed Performance Report

Merging #2772 will degrade performances by 33.78%

_{Comparing desmondcheongzx:local-csv-reader-experiment (7b40f23) with main (cad9168)}

Summary

⚡ 2 improvements ❌ 1 regressions ✅ 13 untouched benchmarks

:warning: Please fix the performance issues or acknowledge them on CodSpeed.

Benchmarks breakdown

	Benchmark	main	desmondcheongzx:local-csv-reader-experiment	Change
❌	test_count[1 Small File]	16.5 ms	24.9 ms	-33.78%
⚡	test_explain[100 Small Files]	52.8 ms	39.9 ms	+32.33%
⚡	test_show[100 Small Files]	597.4 ms	355.8 ms	+67.91%

[desmondcheongzx]

This branch is a little borked. Keeping a record of it but reopening the PR at https://github.com/Eventual-Inc/Daft/pull/3055