The idea is to get a better idea
what the desired formal output of aphid could be.

The likelihood formula used when comparing observed triplet
with a discordant scenario subtly differed from original implementation.

(mostly because `tch` failed to build after system `torch` upgrade)

- Extend calculations by abstracting over `f64` and  `tch::Tensors`.
- The "free parameters" distinction vanishes
  in favour of gradient-tracking declaration.
- Have the calculation leaves step back from `Parameters` to `Scores`,
  ranging freely over the real line.
- Possibly move calculations to CUDA.

- `unresolved_length` (known as `min_d` under C++ implementation)
- `max_clock_ratio`
- `triplet_other_monophyly`

None used/implemented yet.

The first step into calculating likelihood
is to infer the actual triplet topology for any given gene tree.
Take this opportunity to make it unambiguous
what the species topology is in the config file.

Species names and trees names for now, so that:
- Input files can safely be dropped after reading.
- Lifetimes stop infecting the bulk of the code.
- Comparing species names become trivially efficient.

Provide `TreeDataReader` trait to parse nodes/branches data/payloads while
still possibly collecting them for indexing or interning.

Take this opportunity to rewrite the whole filtering procedure
as a plain, neutral "topological analysis", and only filter then
based on the result of this analysis.