Bring typescript benchmark client to parity with rust

[coolreader18]

Description of Changes

• Enable pipelining by default
  • Set defaults for MAX_INFLIGHT_PER_WORKER for spacetime and convex
• Add a warmup period
• Reduce allocations in sdk (not that much of an effect but still improved things)

This gives improved parity with the rust client on my machine.

Expected complexity level and risk

1

Testing

• Yup:

    RESULTS
  ══════════════════════════════════════════════════════════════════════
  
    spacetimedb    ████████████████████████████████████████     80,617 TPS
    convex         █░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░        235 TPS
  
    ╔════════════════════════════════════════════════════════════╗
    ║                                                            ║
    ║       🚀 spacetimedb is 343x FASTER than convex! 🚀        ║
    ║                                                            ║
    ╚════════════════════════════════════════════════════════════╝
  

[Centril]

The code here looks good; I just need to run the benchmarks myself before approving.

[Centril]

This gives pretty much exact parity with the rust client on my machine.

Unfortunately, I cannot confirm this on PhoenixNAP:

With this PR:

All numbers are rounded (ceil) to the nearest whole thousand TPS.

Typescript module

• 110k
• 100k
• 105k
• 108k
• 106k
  average = 105.8k

Rust module

• 108k
• 120k
• 107k
• 117k
• 119k
  average = 114.2k

Without this PR

Typescript module

• 120k
• 120k
• 120k
• 120k
• 119k
  average = 119.8k

Rust module

• 160k
• 162k
• 162k
• 162k
• 160k
  average = 161.2k

[Centril]

(The variance is also clearly higher with a typescript client.)

[Centril]

We decided at a meeting that the numbers were OK but that we would keep around the rust client also.