Keep snapshot PETSc I/O paths compact

[gthyagi]

Summary

Make model snapshots robust when meshes and benchmark output directories have long names on parallel PETSc/HDF5 stacks.

This PR changes only the snapshot backend and its focused tests:

• use deterministic mesh_0000 bulk stems instead of expanding mesh.name into every filename;
• preserve the complete original mesh name/path in wrapper metadata for exact model matching;
• keep snapshot artifacts beside the wrapper, but invoke native PETSc/HDF5 reads and writes from the artifact parent directory using short relative names;
• apply the short-path handling consistently to checkpoint writes, exact model restore, and selective-variable extraction;
• restore the process working directory even when native I/O raises.

Related to #645. This does not change the general Mesh.write_timestep() API or claim to fix long paths outside the snapshot backend.

Root cause

A Zhong thermal-convection checkpoint loaded a reusable mesh whose mesh.name was its 123-character absolute source path. The snapshot backend sanitised that complete name into every PETSc bulk filename, producing a 346-character absolute filename. On Gadi's PETSc 3.25.4 / HDF5 1.12.2p / Open MPI 4.1.7 stack, the first collective write terminated with PETSc signal 11 and exit status 59.

Compact mesh stems fixed that case, but a deliberately longer restart-validation directory still produced 253-271 character absolute paths and reproduced the crash. Passing only basenames to native I/O removes both sources of path growth while preserving the on-disk layout.

Gadi validation

All Zhong runs used the same cellsize=1/16 tetrahedral mesh, 16 MPI ranks, P2/P1/P1 fields, SUPG predictor-corrector, and quadrature degree 4.

State	Job	Longest relevant path	Result
Before fix	178007756	346 chars	Exit 59; 9,552-byte wrapper; empty bulk directory
Compact stems only	178008332	250 chars	Exit 0; complete endpoint checkpoint
Long replay path before relative I/O	178008633	271 chars	Exit 59 at midpoint checkpoint write
Complete fix	178009196	271 chars as an absolute path	Exit 0; exact checkpoint restore and replay

The final job completed in 7:24, used 10.52 GB peak memory, and wrote a 215,179,680 byte checkpoint. Immediate restore and endpoint replay errors were all exactly zero for U, P, T, Tdot, RMS velocity, mean temperature, and upper/lower Nusselt numbers.

Tests

• Gadi focused snapshot suite: 27 passed in 185.12 s.
• Gadi 16-rank Zhong checkpoint/restart/replay: passed with every stored field and diagnostic error equal to 0.0.
• Pre-fix four-rank short-path snapshot control: passed, supporting path length rather than checkpoint state as the failure owner.