Allow custom reqwest::Client or expose HTTP client configuration

[SP-l33t]

The internal reqwest::Client in ClientInner is created with minimal configuration:

let client = ReqwestClient::builder().default_headers(headers).build()?;

This means no TCP_NODELAY, no connection pool tuning, no HTTP/2 window sizing, and no keep-alive configuration. The field is private with no way to inject a custom client.

For latency-sensitive use cases (market making, automated trading), this leaves significant performance on the table:

• No TCP_NODELAY: Nagle's algorithm can add up to ~40ms delay on small packets like order submissions
• Default HTTP/2 stream window (~64KB): Suboptimal for typical market data responses (~469KB), requiring extra round-trips
• No keep-alive-while-idle: Idle connections go stale, causing ~200ms reconnection penalties
• Default connection pool limits: Insufficient for burst order placement (batch cancel + batch place per tick)
• Compression enabled by default: Adds CPU overhead that isn't worth it in colocated/low-latency deployments

Proposed Solution

Either:

  1. Accept a custom reqwest::Client in Config — lets users bring their own tuned client:
  Config {
      custom_client: Option<reqwest::Client>,
      // ...existing fields
  }
  2. Or expose builder options in Config for the most impactful settings:
  Config {
      tcp_nodelay: bool,                          // default: true
      pool_max_idle_per_host: usize,              // default: 10
      pool_idle_timeout: Option<Duration>,         // default: 90s
      http2_initial_stream_window_size: Option<u32>, // default: 524_288 (512KB)
      tcp_keepalive: Option<Duration>,             // default: 30s
      http2_keep_alive_interval: Option<Duration>, // default: 10s
      disable_compression: bool,                   // default: false
      // ...existing fields
  }

Option 1 is simpler and more flexible. Option 2 is more ergonomic for users who don't want to construct a full reqwest client.

Impact

In benchmarking with an optimized reqwest configuration vs the current defaults, the tuned client shows ~15-45ms improvement per HTTP request in colocated environments, which compounds across the multiple API calls needed per trading cycle (get_orders + cancel + place).