ADR 0002 - All tests run on all devices via per-(order, device) entry binaries

rstsr-core's test suite is one device-agnostic body (core_func/, doc_draft/, test_issues/) mod-included by multiple entry binaries - entry_row_cpu.rs, entry_row_faer.rs, entry_col_cpu.rs - each defining its own type DeviceType and static TESTCFG, then pulling in the shared body. The order × device matrix is formed by entry binaries, not by duplicating test code.

Deliberately did not split layout/manipulation tests (results are device-agnostic) from compute tests (device-sensitive) so only the latter run on Faer. Uniformity and Faer tensor-creation path coverage win over compile-time efficiency - every test runs on every device it is compiled for. col_major is compile-time mutually exclusive with row_major (per rstsr-core/Cargo.toml), so row/col are separate binaries by construction.

Device crates reuse the body by symlinking test_utils/ + core_func/ + doc_draft/ and writing their own entry_<device>.rs with a custom TESTCFG skip list. Only DeviceCpuSerial (row-major, entry_row_cpu.rs) is wired in the repo today; entry_row_faer.rs and entry_col_cpu.rs are planned but not yet present. The structure is future-ready for DeviceOpenBLAS and friends.

1. Considered

Layout/compute split per device - rejected (uniformity over efficiency; the body must be symlinkable whole by device crates).

2. Consequences

N binaries compile the same body (slower full builds); per-device skips are runtime via specify_test!/TestCfg, not cfg (avoids recompilation).