Metadata: short strings stored with a backend that say what its data is,
so a reader need not infer it from the file name. Any key may be used;
the ones below are shared between the programs that write and read
backends, and mean the same everywhere.
A preload’s share of the process-wide preload budget, given back when it
is dropped. Keep it next to the preloaded arrays and drop it with them,
so a later preload in the same process sees the memory as free again.
Identifies one of the six 1-D datasets inside a sparse backend.
Used by the streaming write API so we don’t have to add six separate
abstract methods per dtype × (csc|csr) × (data|indices|indptr).
Bytes of (u64, u64, f32) triplets a streaming-write slab may hold, and the
padded size of one such triplet. One definition, because the two streaming
pipelines (the subset trait method and the handlers’ column-selection
writer) each carried their own copy and two memory ceilings drift apart.
Whether loaders preload columns without being asked. On unless
LEGUME_AUTO_PRELOAD is 0, false or off; the preload itself still
answers to preload_within_budget, so data that does not fit in memory
stays on the streaming path.
Whether a preload of nnz entries fits the budget. Unlike
reserve_preload, the reservation is held for the life of the process;
prefer reserve_preload and keep its guard with the preloaded arrays.