fix(block-disk): 4GB pending-write budget default

A single multiturn store burst on a big model is ~1GB of detached
payloads (measured live: dots3-note, 220 blocks at ~14k tokens), which
saturated the old 1GB budget exactly - the burst tail then raced the
sub-second admission wait against the writer's multi-second drain and
lost 2 blocks, poisoning their descendants (4 publish failures) and
collapsing the L2 lineage. The budget bounds transient host RAM for
numpy copies awaiting the background writer, not Metal, and per-turn
bursts are bounded by the delta being stored, so 4GB holds a whole
burst with margin. VMLX_BLOCK_DISK_PENDING_WRITE_BYTES still overrides.