UCT/API: Add strides and counts to uct_ep_put_sgl_zcopy

[michal-shalev]

What?

Add two optional const size_t * parameters, counts and strides, to uct_ep_put_sgl_zcopy (and the matching internal function pointer typedef and dispatch wrapper). When provided, element i represents counts[i] blocks of lengths[i] bytes, each separated by strides[i] bytes, starting at buffers[i] / remote_addrs[i]. When NULL, each element is transferred once (equivalent to count=1, stride=0), preserving the existing behavior.

Why?

The current SGL zero-copy put requires one entry per transferred buffer, so a regular strided layout has to be expanded into long buffers/lengths/remote_addrs arrays. Allowing per-element counts/strides lets the caller describe such patterns compactly.

How?

API-only change in src/uct/api/v2/uct_v2.h, with the matching typedef in src/uct/base/uct_iface.h and dispatch wrapper in src/uct/base/uct_iface.c updated. Implementation will follow in subsequent PRs.

[bosilca]

This is awesome, I've been waiting for such an extension for a long time. Lately I discovered your devel branches, and I've been using them [almost] successfully.

Would it be possible to provide a way to have each one of the rkeys, counts and strides arrays as a single value, with the understanding that in such a case all elements have the same value ?