Remove the serialization size bottleneck

[FTRobbin]

• Replace serde_json with serde/flexbuffers.
• Implement a debug feature that prints the size of serialized egraphs.

Output example:

cargo run --bin poach tests/taylor51.egg tempo/ size-report

egglog::EGraph : 4.97GB
. rulesets : 4.06GB (81.82%)
. functions : 0.44GB (8.95%)
. backend : 13.59MB (0.27%)
. type_info : 0.24MB (0.00%)
. overall_run_report : 148.99KB (0.00%)
. proof_state : 78B (0.00%)
. schedulers : 27B (0.00%)
. commands : 7B (0.00%)
. command_macros : 7B (0.00%)
. pushed_egraph : 3B (0.00%)
[1/1] taylor51 : SUCCESS
0 failures out of 1 files

*Numbers were off by a bit from the Mar 2nd meeting version because I eyeballed the size from the number in bytes for the meeting and used base 1000 instead of 1024.

• Investigate the problem in rulesets.

The problem is in span. The egglog frontend uses spans throughout its ASTs for debugging.

pub enum Span {
    Panic,
    Egglog(Arc<EgglogSpan>),
    Rust(Arc<RustSpan>),
}

...

pub struct EgglogSpan {
    pub file: Arc<SrcFile>,
    pub i: usize,
    pub j: usize,
}

However, the default implementation generated by #[derive(serde::Serialize)] performs a deep copy for Arc<SrcFile>, which means every node in the AST would carry a copy of the whole input egglog program, causing a significant blowup that scales quadratically with the input file size (# AST nodes * filesize).

• Remove the span serialization bottleneck.

I implemented a quick hack that serializes span into unit. For the example tested above, we now have a 300x overall reduction and a 1600x reduction for rulesets and functions:

&egglog::EGraph : 17.12MB
  backend : 13.56MB (79.24%)
  rulesets : 2.53MB (14.80%)
  functions : 0.72MB (4.25%)
  type_info : 0.24MB (1.42%)
  overall_run_report : 148.99KB (0.85%)
  proof_state : 78B (0.00%)
  schedulers : 27B (0.00%)
  pushed_egraph : 3B (0.00%)
[1/1] taylor51 : SUCCESS
0 failures out of 1 files

• Add experiments previously blocked by the serialization size.

Run extract runmode on easteregg, herbie-hamming, herbie-math-rewrite, herbie-math-taylor benchmarks.

Fixed minor bugs.

• Update nightly frontend to present the new data

Added radio buttons to pick a benchmark.

Added a new chart to show the speedup of using Poach vs. Vanilla.

Added a checkbox for including deserialization time in Poach time.

• Get a nightly run. Link

• List what research questions to ask and how to look at the data.

• Add file size statistics to the nightly report. (Link in the #poach channel)

• Is the serialization size linear to the egraph size? - Yes (~500bytes per tuple)

• Is the serialization/deserialization time linear to the serialization size? - Yes

• Clean up hacks.

• Get another nightly run. Link Results reproduced.

• Investigate if egglog's hashcon works as I thought it does