Use OpenDP accountant for budget composition and look into decimal or rational

[tholop]

We have a Budget trait to define filter capacity and requests. It's currently only a marker trait, and we can keep it this way. The marker trait is implemented for our PureDP budget struct and the corresponding filters. We should change the struct to use OpenDP's MaxDivergence, then call the composition function inside the filter: https://docs.rs/opendp/0.12.0-nightly.20250205.1/src/opendp/combinators/sequential_compositor/noninteractive/mod.rs.html#111-118. I haven't checked in detail if this is enough to obtain a filter and whether this will require to import all sorts of OpenDP traits, but it seems quite doable.

We can even be a bit more general and implement a generic filter for any type implementing https://docs.rs/opendp/0.12.0-nightly.20250205.1/opendp/combinators/trait.BasicCompositionMeasure.html, which could give us RDP and zCDP for free, but needs a bit more investigation. We'd also need to match the mechanism/noise scale to the appropriate measure, e.g. use only PureDP for Laplace noise. See #10 (non blocking, either issue can be tackled first).

[tholop]

EDIT: Even though the budget in OpenDP is f64, Michael told me the operations for filters were carefully implemented to take floating point rounding errors into account. It would be good to find the exact implementation details in the OpenDP Rust codebase though.

Since OpenDP seems to be using f64, we might instead (or in addition, if we still implement an OpenDP trait) want to implement our own exact accounting using fixed point or rational numbers. Ideally with a lightweight library that is compatible with our Firefox integration, so that might rule out certain Bigint libraries. Some candidates:

• https://docs.rs/num-rational/latest/num_rational/type.Rational64.html
• https://docs.rs/rust_decimal/latest/rust_decimal/
• https://docs.rs/fastnum/latest/fastnum/index.html
• https://crates.io/crates/rational

But overall I think that this issue is pretty low priority compared to developing the rest of pdslib, as long as pdslib is not used in a production environment.