test: Eliminate redundant work from test runs

[zharinov]

A warm make test took about 101 seconds even though the same workspace completed near 32 seconds with shared-process execution. The gap came from work that Nextest's process-per-test model amplified: grammar fixtures launched cargo metadata at runtime, all-language invariants rebuilt the same grammars in separate processes, and two generated-Rust suites launched competing rustc jobs.

This change removes that redundant work while keeping Nextest and its per-test isolation:

• Resolve the exact Arborium direct dependencies once in plotnik-tests/build.rs and embed their grammar JSON in test artifacts. A representative isolated VM fixture fell from 233 ms to 62.6 ms.
• Run the structure-table and ABI invariants in one process so the language registry's OnceLock caches are effective. Their failures are still aggregated, and both former filter names continue to select the combined test.
• Compile the committed 04-emit Rust sections and the 06-vm generated matchers in one trybuild program. Conformance helpers live under a unique crate-root module so they cannot accidentally satisfy imports emitted by a broken generator.

On the same checkout, warmed full-suite runs fell from 101.0 seconds to 34.5–39.0 seconds, with all 1,606 Nextest trials passing. A shared-process comparison completed in 31.6 seconds, leaving too little runner-specific upside to justify giving up Nextest's isolation and reporting.