[ashhart]

Part of #2003. Stacked on #1928 (@jschmied) and #1887: the first two commits are #1928's own, and the series sits on #1887's TP page-drop deferral. It lands after both, and #1928's commits drop out then.

#1928's read-ahead (TF_LOAD_PREFETCH: reader threads read a layer's checkpoint intervals while its loader packs the weights) is ported to the other CUDA loaders. On these families it's opt-in (unset is 0) until each one's GB10 load receipts are in. GLM keeps its existing default, and #1928's own two families are unchanged.

• TF_LOAD_PREFETCH=0 is each family's main path byte for byte: no thread, no read, no change to release order.
• Prefetch only reads. Uploads and page drops keep main's order, and TP ranks never drop pages a peer still reads (CUDA: Qwen4 and tensor-parallel loaders drop the checkpoint's pages after every rank is ready #1887's deferral).
• Host tests per family: off starts no reader; on reads each layer once, in order; a failed read refuses by name and frees everything.

Reviewed independently three times. The host suite, the Linux CUDA build and lean (424) pass. GB10 load-time cells per family come before any default changes.