Fix/qdma partial transfer

[amd-vserbu]

Summary

Fix QDMA exact-length transfers by clipping the final scatter-gather entry per transfer instead of requiring callers to pad all transfer lengths to 4 KiB page multiples.

• Keep BUF_CREATE buffers page-multiple and keep the reusable per-buffer SGL unchanged.
• Add a per-transfer clipped SGL path for requests whose byte count ends inside the final page, while preserving the existing cached-SGL fast path for page-multiple transfers.
• Update the QDMA ABI documentation and UAPI comments to describe exact byte-count transfers, with a note that non-4 KiB-multiple transfers are not speed-optimized and should currently be reserved for PDI-style exact byte streams.
• Add regression coverage for partial-length transfers in the driver kselftest and libslash wrapper tests.

Motivation

The design-writer path needs to send exact PDI byte counts. Padding a PDI stream with trailing zeroes is not something we should rely on as a general programming contract, but the QDMA ioctl previously documented page-multiple transfer lengths and effectively forced callers toward padding.

This change lets vrtd continue submitting the real PDI length while keeping normal application DMA buffers on the optimized 4 KiB-multiple path.

Draft

This PR is a draft now as it still needs to be tested on hardware.

[JOOpdenhoevel]

I've tested the changes in hardware with the base design from commit 4c79427 (Gitlab CI pipeline 133074), and it works without issues.

[taraxacum45e9a]

Tested examples 00 and 02 on this branch with the base design built from 4c79427. Both examples run successfully on hardware.