HPU backend fixes

[Bapt-Roux]

Summary

• Fix mmap ptr lifetime — replaced explicit munmap call with a Drop trait implementation, ensuring correct memory cleanup via Rust's drop glue
• Fix HpuDevice init — updated missed function calls left behind from a prior HPU backend refactor
• Fix DMA queue write guard — now reads and validates the max queue value before attempting to write it, preventing out-of-bounds writes
• Fix DMA queue PATH — added missing PATH entry required to update DMA queues

Chores

• Updated psi64.hpu with latest bitstream and firmware from main
• Extended the HPU bench GitHub workflow to run on a broader set of HPU targets

[pgardratzama]

bench from github workflow are running and manual tests look good

[github-actions]

✅ Backward-compat snapshot: everything looks good! No backward-compatibility issues detected.