fix: harden table, vector, and runtime tests

[belowzeroff]

Summary

This PR fixes three small user-visible/runtime issues in the Rust bindings:

• Normalize Table::head and Table::tail counts by magnitude, matching the table transform docs.
• Return &[] for empty zero-copy vector slices before reading the core data pointer.
• Run CI tests with RUST_TEST_THREADS=1, matching the single process runtime constraint documented by the test suite.

User-visible bugs

head(-n) and tail(-n) returned the wrong side

For a table whose size column is [10, 20, 30, 40, 50], the docs say head and tail use the magnitude of n, while take uses the sign.

Before:

head(-2) -> [40, 50]
tail(-2) -> [10, 20]
take(-2) -> [40, 50]

After:

head(-2) -> [10, 20]
tail(-2) -> [40, 50]
take(-2) -> [40, 50]

This also handles i64::MIN without overflowing.

Empty typed slices could read an invalid FFI data pointer

Rust requires slice::from_raw_parts(ptr, 0) to receive a non-null, aligned pointer even for an empty slice. The binding previously read raw::data(...) before constructing an empty slice, which made empty core vectors depend on the exact pointer returned by the C core.

The binding now returns Ok(&[]) before reading the data pointer when len == 0.

CI could run runtime tests in parallel

The runtime is process-global and tests already document serialized execution. CI still ran cargo test --workspace without RUST_TEST_THREADS=1, allowing integration tests to race runtime creation and fail with an already-live runtime.

The test step now sets RUST_TEST_THREADS=1.

Tests

• cargo fmt --all -- --check
• cargo clippy --workspace --all-targets -- -D warnings
• env RUST_TEST_THREADS=1 cargo test --workspace