Redesign `CosineEmbeddingLoss` to fit the unified loss compute path

[MoonFlowww]

Motivation

The unified loss dispatch path in Model calls compute(descriptor, prediction, target, weight) with a single prediction tensor. CosineEmbeddingLoss fundamentally needs two separate embedding inputs (x1, x2). Right now callers must pack both into a stacked tensor shaped [B, 2, D…] or [2, B, D…] — an undocumented, fragile convention that nothing in the public API communicates.

This makes it impossible to use CosineEmbeddingLoss through the standard Model::train() path without prior knowledge of the packing trick, breaking the "safe API" promise.

Current State

src/loss/details/cosine_embedding.hpp:9 carries:

/// TODO: rework
/// - using single path : std::visit([&](const auto& d){ return Loss::Details::compute(d, prediction, target, weight); }, *loss_descriptor_);
///      --> Cos Embedd expect 2 inputs

The current workaround silently accepts prediction.size(1) == 2 or prediction.size(0) == 2 and splits — no compile-time or runtime API contract enforces this.

Proposed Change

Two viable approaches:

Option A — optional second input slot in the unified compute path:

torch::Tensor compute(const CosineEmbeddingDescriptor& d,
                      const torch::Tensor& prediction,
                      const torch::Tensor& target,
                      const std::optional<torch::Tensor>& weight = std::nullopt,
                      const std::optional<torch::Tensor>& input2  = std::nullopt);

input2 is only consulted by CosineEmbedding; all other losses ignore it.

Option B — PairTensor wrapper:

struct PairTensor { torch::Tensor x1, x2; };

Callers construct a PairTensor explicitly. The loss dispatcher checks via if constexpr whether the descriptor requires a pair and unpacks accordingly.

Option B is cleaner long-term (no proliferating optional args) and makes the requirement visible at the call site.

Acceptance Criteria

• CosineEmbeddingLoss usable through the standard Model::train() path without manual tensor packing
• The packing convention (if kept) is enforced and documented at the API boundary
• /// TODO: rework comment removed
• Existing behaviour (margin, reduction modes) preserved
• No other loss functions affected