Per-row drain in binary protocol + memory issue in text path

[Selerski]

@barakalon Hey, I've used v2.5.5 in an internal system, since then you've improved performance with PR 51 (more efficient buffer drain) and PR 71 - fix O(n²) encoding.

I found a few more small bottlenecks below for prepared statements. I have a local branch with fixes but no write access, if you're interested please grant me write access so I can submit a PR for your review

PERFORMANCE ANALYSIS

1. Binary protocol still flushes after every row
2. Text protocol buffers the whole resultset in memory before sending anything
3. Inbound packets hog memory - Inbound packets over 16 MB arrive as chunks and are combined with bytes +=
4. **read_str_null - same issue as 3

LOCAL BENCHMARK

Binary protocol:

Build	Before	After
pure Python	88k rows/s	250k rows/s (×2.8)
mypyc	112k rows/s	321k rows/s (×2.9)

Flushes for a 10k-row query: 10,010 → 20.

Text protocol:

×1.5 on pure Python (214k → 315k rows/s), no change on mypyc

Quadratic loops:

• 1 GB inbound packet: 22.3 s → 0.54 s
• 1 MB null-terminated string: 15.3 s → 0.08 s

[barakalon]

This sounds very promising!
You shouldn't need write access to open a PR - just create a fork.
Looking forward to reviewing.