feat(fuse): stream large sparse writes

[Kazimierzsier]

Summary

Add a bounded-memory write path for large and sparse FUSE objects without increasing the 1 GiB in-memory threshold. Dirty contents spill to a sparse temporary file, stream through the worker data plane, and upload to S3, GCS, or Azure without materializing the complete object in memory.

This PR is stacked on #346. Its diff will shrink to the streaming-write commit after the prerequisite POSIX stack merges.

Closes #347.

Changes

• Keep objects at or below 1 GiB in the existing in-memory working copy.
• Promote writes beyond the threshold to a sparse NamedTempFile.
• Enforce a separate 4 GiB logical object limit so large offsets remain bounded.
• Stream staged files from FUSE to workers with a size-aware timeout.
• Receive large worker PUT bodies in fixed 8 MiB chunks and preserve zero ranges as sparse holes.
• Add BackendStore::put_file and streamed HTTP request support.
• Hash S3 payloads with a fixed 1 MiB buffer before SigV4 signing.
• Send GCS and Azure streamed PUTs with an exact Content-Length.
• Avoid inserting streamed multi-block objects into the single-block worker cache.
• Cover sparse staging, worker transfer, backend request construction, and real-kernel FUSE writeback.

Large truncate remains memory-backed and retains the existing 1 GiB limit. This PR addresses large sparse writes; it does not claim large truncate support.

Test plan

Implementation commit: 7fd6bbf2725dd80ebbd33a6f0dfa6fe1018a2e6f

pjdfstest runner commit: 26f10147a214ed8dc6a59298aca4c5c3a4011263

• cargo test -p talon-fuse --lib --features mount: 122 passed
• cargo test -p talon-backend --lib: 77 passed
• cargo test -p talon-worker --lib: 89 passed
• cargo test -p talon-fuse --test mount_e2e --features mount --no-run
• cargo check --workspace --all-targets
• Strict clippy for core, backend, worker, and FUSE mount targets
• cargo fmt --all -- --check
• git diff --check
• Exact implementation SHA passed 15/15 privileged real-kernel mount tests in falcon-phx-ca
• open/25.t: 6/6 passed through a real kernel FUSE mount
• open,rename,unlink,truncate,mkdir,rmdir: 5,981/5,981 passed

The open/25.t validation writes one byte at offset 2 GiB + 1, verifies the resulting size, reads the sparse hole, and reads the written byte. The FUSE layer, worker, and pjdfstest mock all use bounded-memory streaming for this path.

Checklist

• PR title follows Conventional Commits
• Public behavior is covered by tests
• The 1 GiB in-memory threshold is unchanged
• All code, tests, commit messages, and PR text are in English

[beinan]

Superseded by #351, which cherry-picks this commit onto current main with your authorship preserved and the content unchanged.

The conflict was historical rather than semantic: the lower half of this stack was squash-merged (#309 → #344), which rewrites history, so this PR still carried commits whose content was already upstream under different SHAs. Cherry-picking the top commit applies cleanly with no conflict markers, which is what confirmed that.