Struct std::TensorArena

Every op allocates its output. In a training loop that is a malloc (and, for fresh memory, page faults) per op per step: on a small MLP that cost more than the arithmetic. A TensorArena is a chunked bump allocator for a step's temporaries: between a.begin() and tensor_arena_end() every tensor created on this thread (data, shape, gradient) comes from it, and a.reset() drops them all at once while keeping the memory -- already faulted in -- for the next step.

struct TensorArena a = TensorArena::new(0UL);
while (step < steps) {
    a.begin();
    ... forward, Tensor::backward, optimizer update ...
    tensor_arena_end();
    a.reset();     // this step's tensors are gone
}
a.free();

Tensors created outside the arena (parameters) stay on the heap, and so do their gradients, so they survive resets. Tensor::free() on an arena tensor does nothing. A tensor must not be used after its arena is reset.