Harden pread and pwrite short I/O handling

[fallintoplace]

Summary

Harden the POSIX pread/pwrite paths in CuFileDriver against EINTR, short transfers, EOF, and error propagation.

What changed

• add centralized retrying helpers for pread/pwrite
• add shared buffered read/write helpers that stop cleanly on EOF or partial transfer instead of assuming the requested byte count always completed
• update the internal POSIX and O_DIRECT buffered/page/block paths to copy only the bytes that were actually transferred
• add a regression test for reading past EOF that verifies unread bytes in the destination buffer stay untouched

Why

Several loops in cufile_driver.cpp assumed that each pread/pwrite returned the full requested size.

When that assumption breaks, the old code could:

• subtract a negative count from an unsigned remaining size
• copy more bytes to host/device buffers than were actually read
• keep advancing totals as if a short transfer had succeeded
• get stuck or report misleading success counts

This change keeps the public behavior close to POSIX short-read semantics while making the internal buffered/page-aligned logic robust against partial transfers and interruptions.

Validation

• reviewed and refactored the internal POSIX/O_DIRECT read/write paths in cufile_driver.cpp
• added a regression case in test_cufile.cpp for a short read near EOF
• ran git diff --check
• attempted local CMake configure, but this machine cannot build the project because FindCUDAToolkit fails to locate nvcc

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.