RFC: dp accounting should use rust as its backend

[amyssnippet]

I am proposing an optional, high-performance backend for dp-accounting that resolves the computation bottleneck in RDP accounting.

By offloading the exact RDP math (Poisson Subsampled Gaussian) to a vectorized Rust extension, I achieved a 40x-45x speedup on standard workloads while maintaining floating-point accuracy ($10^{-12}$

Proposed Solution

I have developed dp-accelerator, a drop-in Rust replacement that:

1. Implements the exact _compute_log_a_frac logic from Google's dp_accounting using statrs.
2. Vectorizes computation across all $\alpha$ orders.
3. Exposes a safe Python interface via PyO3.

This package is already published on PyPI (pip install dp-accelerator) and uses Apache 2.0.

Benchmarks (Intel Mac i5 8th gen)

Config	Steps	Python (ms)	Rust (ms)	Speedup
MNIST	1k	36.5	0.86	42x
ImageNet	90k	30.0	0.69	43x
Batch (10 checks)	1k-60k	333.4	8.8	37x

Integration Plan

I have a working branch where this is implemented as a soft dependency:

• dp-accounting does not need to compile Rust.
• If dp-accelerator is installed, analysis.py detects it and routes the computation to the fast path.
• If not installed, it falls back to the existing Python implementation with an alert kind of warning to suggest using the dp-accelarator package for faster

I have prepared a technical report (submitted to techrxiv) detailed in the repo: https://github.com/AxiomaticLabs/dp-accelerator

Would the team be open to a PR integrating this optional backend?