llama-quant : overlap compute and write with double buffering

[nuri-yoo]

Overview

Pipeline the main quantization loop so that the previous tensor's disk write overlaps with the current tensor's compute. Uses a background writer thread and double-buffered work areas.

On macOS (where mmap is disabled) this avoids blocking the main thread on sequential writes. Output is bit-identical to the original.

Benchmark (Qwen3-1.7B Q5_K_M → Q4_0, M4 Pro, macOS):

	Wall time
Before	~1.86s
After	~1.30s

The improvement should scale with model size since larger tensors have proportionally more I/O to overlap with computation.

Testing

• Output is bit-identical (MD5 match) with the original
• --dry-run mode works correctly
• Exception safety: requantize error exits cleanly without crash
• test-quantize-fns passes
• test-quantize-perf passes

Additional information

Closes #20829

Requirements

• I have read and agree with the contributing guidelines
• AI usage disclosure: NO