feat: add fetch/unfetch to Vfs trait (xFetch/xUnfetch, iVersion 3)

[russellromney]

Context

Follows from #79 and #80. PR #80 fixes the SEGFAULT by setting iVersion=2. This PR is the alternative: properly implement iVersion=3 by adding fetch/unfetch to the Vfs trait.

Changes

Add fetch() and unfetch() methods to the Vfs trait:

fn fetch(&self, handle: &mut Self::Handle, offset: i64, amt: usize)
    -> VfsResult<Option<NonNull<u8>>> { Ok(None) }

fn unfetch(&self, handle: &mut Self::Handle, offset: i64, ptr: *mut u8)
    -> VfsResult<()> { Ok(()) }

Default implementations decline mmap (fetch returns None, unfetch is a no-op). SQLite gracefully falls back to xRead. VFS implementations that want memory-mapped page reads can override fetch() to return NonNull pointers.

C shims x_fetch and x_unfetch wired into io_methods. iVersion stays at 3, now legitimately.

Use case

We're building turbolite, a SQLite VFS that serves queries from S3 with a local NVMe cache. With fetch(), we can mmap the cache file and return direct pointers to SQLite, eliminating pread syscalls for warm-cache reads (~5us to <1us per page).

Tests

4 new tests in tests/fetch_test.rs:

• Basic write/read roundtrip with default fetch (decline mmap)
• Concurrent WAL (1 writer + 4 readers, 3 seconds) with default fetch, 30/30 pass
• Checkpoint under load (2500 rows, explicit TRUNCATE checkpoint)
• iVersion=3 end-to-end validation

Relationship to #80

Either PR fixes the SEGFAULT. #80 is the minimal fix (iVersion 3 -> 2). This PR is the proper fix (implement the methods that iVersion 3 declares). They're alternatives; merging either one works. If you prefer the minimal approach, merge #80 and this becomes a follow-up feature.

[russellromney]

simplifying tests.

[russellromney]

Ready for review.

[carlsverre]

The implementation looks good. Can you add fetch/unfetch support to your minimal vfs to exercise it? Make sure that something asserts they are actually reached in testing. I'm not convinced from reading the docs that they are. It sounds like mmap_size has to be set?

[russellromney]

Good point. Adding

[russellromney]

updated. test VFS now implements real mmap fetch/unfetch, and an atomic counter asserts that SQLite actually calls fetch() (requires PRAGMA mmap_size). ready for review.

[carlsverre]

just some small tweaks then we can land this! thank you

[russellromney]

All addressed, great feedback thx.