Feature: Quantized Inference Kernels

[ShlokVFX]

Implement quantized inference kernels for linear and attention layers.

This includes int8 or mixed-precision kernels with efficient
dequantization and accumulation paths.

Planned Benchmarks

• Latency vs precision
• Accuracy comparison
• Memory footprint reduction

Learning Objectives

• Quantization schemes
• Dequantization strategies
• Performance vs accuracy tradeoffs