feat(engine): named test definitions — define a check once, apply by name across models

[hugocorreia90]

What

Reusable named data-quality tests — the declarative analogue of dbt's generic tests (C3). Define an assertion once in models/test_definitions.toml and apply it by name across many models/columns, instead of inlining the same [[tests]] block everywhere.

# models/test_definitions.toml
[positive_amount]
type = "expression"
expression = "amount > 0"

[known_status]
type = "accepted_values"
column = "status"            # default column
values = ["pending", "shipped", "delivered"]

# models/fct_orders.toml
[[tests]]                    # inline tests still work
type = "unique"
column = "order_id"

[[use_test]]                 # apply a named definition
name = "positive_amount"
column = "amount"

[[use_test]]
name = "known_status"        # uses the definition's default column
severity = "warning"

Design

• References resolve into ordinary TestDecls at load and append to inline [[tests]], so they execute identically through rocky test — no change to the test-execution path.
• An unknown [[use_test]] name fails the load (ModelError::UnknownTest), so a typo can't silently drop a check.
• Loading rides a side-loader (load_test_definitions_from_dir). Resolution generalizes the config-group threading (from feat(engine): config groups — one definition fans out routing + strategy to N models #917) — the groups-only param becomes a shared ModelLoadContext { groups, test_defs }; *_with_groups → *_with_context; bare load_model_pair/parse_model_inline pass an empty context. Zero ModelConfig churn, no output-schema change (verified: no codegen/schema drift).
• Wired through both loaders the compiler uses — .sql (load_models_from_dir) and .rocky (salsa) — each with a regression test, and verified end-to-end through rocky compile (valid references resolve; unknown ones error).

Tests

rocky-core: applied-by-reference (default column + severity), column-override + inline coexistence, unknown-name error. rocky-compiler: .rocky model resolves a named test through the salsa loader. Full suites green (1357 core + 325 compiler).

Documented in concepts/data-quality-checks.md.

cargo fmt + clippy --all-targets clean; no codegen/schema drift.