Add a universal CSV/JSON loader (`Load::Universal`)

[MoonFlowww]

Motivation

Nott's stated goal is to be easier than raw LibTorch. Today, users with any custom tabular dataset (CSV or JSON) must implement their own parsing, tensor construction, and train/test splitting — the same boilerplate PyTorch users write every time. The infrastructure for this is already partially in place (Data::Type::CSV, CSVParameters, column selection), but no loader actually reads a file and produces tensors.

A universal loader removes the single biggest friction point for new users trying Nott on their own data.

Current State

src/data/load/types.hpp defines:

struct CSVParameters {
    bool has_header = true;
    char delimiter = ',';
    std::vector<std::string> columns; // empty = all columns
};
struct CSV { std::string file; CSVParameters parameters{}; };

src/data/load/load.hpp has CSV parsing helpers (split_csv_line, trim_copy) but no public loader function that returns tensors. Users have no supported path to use these types.

Proposed Change

Add Load::Universal in src/data/load/load.hpp:

namespace Nott::Data::Load {

struct UniversalDescriptor {
    std::variant<Data::Type::CSV /*, Data::Type::JSON */> source;
    std::vector<std::string> input_columns;  // empty = all except target_column
    std::string              target_column;
    Data::Type::GlobalParameters split{};    // train/test fraction + shuffle
};

struct UniversalDataset {
    torch::Tensor train_inputs;
    torch::Tensor train_targets;
    torch::Tensor test_inputs;
    torch::Tensor test_targets;
};

UniversalDataset Universal(const UniversalDescriptor& descriptor);

} // namespace Nott::Data::Load

Implementation reuses the existing split_csv_line / trim_copy helpers and GlobalParameters splitting logic already used by MNIST/CIFAR loaders. JSON support can follow in a subsequent PR using the same descriptor pattern.

Usage target:

auto dataset = Nott::Data::Load::Universal({
    .source         = Nott::Data::Type::CSV{"my_data.csv"},
    .input_columns  = {"open", "high", "low", "volume"},
    .target_column  = "label",
    .split          = {.train_fraction = 0.8f}
});
model.train(dataset.train_inputs, dataset.train_targets, ...);

Acceptance Criteria

• Load::Universal accepts a Data::Type::CSV descriptor and returns UniversalDataset
• input_columns empty → all columns except target_column used as inputs
• Train/test split and shuffle respect GlobalParameters
• Missing or misnamed columns produce a clear std::invalid_argument
• Consistent return type with existing MNIST/CIFAR loaders
• No dependency added beyond what load.hpp already includes