Replace hardcoded `reserve(8)` with dataset-size-aware reservation in calibration streaming

[MoonFlowww]

Motivation

Vector reallocation during the calibration streaming loop is a latency spike that grows with dataset size. The current hard-coded reserve(8) assumes at most 8 batches — correct for toy datasets, wrong for any real evaluation run. On a 100k-sample dataset with batch size 32 that is ~3125 batches and ~11 reallocations, each doubling the allocation and copying all accumulated logit chunks.

Current State

src/core.hpp:2258–2261:

std::vector<torch::Tensor> logits_chunks;
std::vector<torch::Tensor> target_chunks;
logits_chunks.reserve(8); // TODO: Modify
target_chunks.reserve(8);

Both evaluation_inputs.size(0) (total samples) and streaming_options.batch_size are available at this point in the function, so the correct capacity is computable for free.

Proposed Change

const auto n_samples       = evaluation_inputs.size(0);
const auto batch_sz        = static_cast<int64_t>(streaming_options.batch_size);
const auto estimated_batches = (n_samples + batch_sz - 1) / batch_sz;

std::vector<torch::Tensor> logits_chunks;
std::vector<torch::Tensor> target_chunks;
logits_chunks.reserve(static_cast<std::size_t>(estimated_batches));
target_chunks.reserve(static_cast<std::size_t>(estimated_batches));

No behavioural change — purely an allocation optimisation.

Acceptance Criteria

• reserve() size derived from actual input size and batch size
• Zero reallocations during a standard streaming evaluation loop
• // TODO: Modify comment removed
• No change to outputs, metrics, or evaluation semantics