Reuse CUDA quant output buffers instead of fresh allocation

[SaschaOnTour]

Problem / Motivation

In cuda/quantize.rs:116-117, both packed_flat and scales_flat tensors are freshly allocated on every quantize call. Same issue as #33 but for the quantization path.

Solution

Pre-allocate scratch buffers for quant output, reuse across calls.

Key files

• turboquant/src/cache/cuda/quantize.rs:116-117 — current fresh allocations

Acceptance criteria

• No Tensor::zeros in the quant hot path
• Scratch buffers allocated once, reused
• All tests pass: cargo nextest run --features cuda
• cargo fmt --check clean